- Open Access
Mechanism-anchored profiling derived from epigenetic networks predicts outcome in acute lymphoblastic leukemia
© Yang et al; licensee BioMed Central Ltd. 2009
- Published: 17 September 2009
Current outcome predictors based on "molecular profiling" rely on gene lists selected without consideration for their molecular mechanisms. This study was designed to demonstrate that we could learn about genes related to a specific mechanism and further use this knowledge to predict outcome in patients – a paradigm shift towards accurate "mechanism-anchored profiling". We propose a novel algorithm, PGnet, which predicts a tripartite mechanism-anchored network associated to epigenetic regulation consisting of phenotypes, genes and mechanisms. Genes termed as GEMs in this network meet all of the following criteria: (i) they are co-expressed with genes known to be involved in the biological mechanism of interest, (ii) they are also differentially expressed between distinct phenotypes relevant to the study, and (iii) as a biomodule, genes correlate with both the mechanism and the phenotype.
This proof-of-concept study, which focuses on epigenetic mechanisms, was conducted in a well-studied set of 132 acute lymphoblastic leukemia (ALL) microarrays annotated with nine distinct phenotypes and three measures of response to therapy. We used established parametric and non parametric statistics to derive the PGnet tripartite network that consisted of 10 phenotypes and 33 significant clusters of GEMs comprising 535 distinct genes. The significance of PGnet was estimated from empirical p-values, and a robust subnetwork derived from ALL outcome data was produced by repeated random sampling. The evaluation of derived robust network to predict outcome (relapse of ALL) was significant (p = 3%), using one hundred three-fold cross-validations and the shrunken centroids classifier.
To our knowledge, this is the first method predicting co-expression networks of genes associated with epigenetic mechanisms and to demonstrate its inherent capability to predict therapeutic outcome. This PGnet approach can be applied to any regulatory mechanisms including transcriptional or microRNA regulation in order to derive predictive molecular profiles that are mechanistically anchored. The implementation of PGnet in R is freely available at http://Lussierlab.org/publication/PGnet.
- Acute Lymphoblastic Leukemia
- Mutual Information
- Seed Gene
- Reverse Engineering Method
- Leukemia Phenotype
By design, predictors of outcome based on gene expression profiles are based on gene lists that do not require knowledge of biological processes or molecular mechanisms . Though expression arrays have been widely studied to improve prediction of clinical outcome and to aid the decision of treatment strategy for cancer, the resulting long list of genes lacking mechanistic background is thus difficult to interpret to infer their biological or clinical implications. Additionally, a poor outcome may be caused by a diversity of molecular disorders, for which the individual contribution may vary in different patients suffering from the same cancer . In some cases, profiles are accompanied with follow-on enrichment studies or curated annotations that predict their possible mechanisms; while in other cases, functional clustering has been proposed to understand microarray data profiles . In this manuscript, we propose a novel computational strategy based on genes associated to known biological mechanisms to derive mechanism-anchored expression profiles ab initio that can accurately predict disease outcome.
We hypothesized that those co-expression modules, which are predictive of outcome, can be computationally derived from genes known to regulate or to be regulated by epigenetic mechanisms in previous studies and from novel microarray expression specifically designed for a new phenotype for which the epigenetic mechanisms may not be well understood [4, 5]. Nearly every cancer consists of genetic mutations of the transformed cells as well as epigenetic abnormalities of non-mutational changes to DNA that lead to alterations in gene expression . While genetic abnormalities found in cancer typically affect cancer-promoting oncogenes and tumor suppressor genes, the epigenetic regulation of molecular functions involves reversible interactions which can affect gene expression such as (i) DNA methylation , (ii) histone modification , (iii) RNA transcription and the resulting proteins [9, 10] or miRNAs , that influence chromatin structure. For example, histone deacetylation and the methylation of the promoter region can affect binding of transcriptional factors to these DNA regions and result in transcriptional silencing partly due to chromatin remodeling . Indeed, combination therapy with inhibitors of DNA methyltransferase and histone deacetylase is under investigation in cancer [13–15]. Additionally, epigenetic events occur in the coordinated behavior of epigenetic proteins that regulate gene expressions . To demonstrate the applicability of the proposed p henotype-g enotype-net work" method (PGnet), a set of known biological mechanism-anchored genes are required as the "seed" (input). Although this method can be generalized to other molecular mechanisms and other diseases, we focused this study on epigenetic alterations in acute lymphoblastic leukemia (ALL).
To compute co-expression modules of genes in disease and to infer their interplay generally require the integration of data from a wide variety of sources [16, 17]. For example, some computational methods have been developed to indentify shared regulatory inputs, functional pathways and genetic interactions [2, 18–21]. We have also previously shown that co-expression patterns of genes found in expression arrays designed around specific phenotypes can be recapitulated in gene-phenotype relationships derived from database/literature mining . Further, genome-scale reverse engineering of regulatory mechanisms in expression arrays have been developed and successfully applied in mammalian cells . For example, A method called ARACNE has been shown to be effective in practice by using a mutual information theoretic approach which focuses on direct co-expression of genes [24, 25]. However, unsupervised combinations of every molecular element that may interact via one or more intermediaries can lead to a problem of multiplicity due to the escalating number of comparisons and thus to a loss of statistical power. Another method, FunNet, addresses multiplicity by combining gene expression data with Gene Ontology [26, 27] or KEGG  annotations and further performs transcriptional functional analysis over co-expression . Another method, StAM, identifies expression signature by focusing on biological processes which can characterize subgroup of patients . However, these methods are not designed to compute regulatory networks that would also be differentially expressed in multiple phenotypic contexts as well as co-expressed in each individual. Our approach differs from these previous methods in that genes are integrated to the profile signature if: i) they are associated ab initio to the biological mechanism of interest (here epigenetics), and ii) they are derived from the a non-parametric statistic taking into account comprehensive expression patterns of the every gene in the microarray rather than from a subset of the differential-expressed ones.
We hypothesize that using supervised pair-wise measurements from microarray data together with robust feature selection technology , we are more likely to construct meaningful, epigenetic mechanism-anchored, co-expression networks that are predictive of leukemia outcome. To this end, we propose a novel supervised non-parametric algorithm (PGnet) that builds a tripartite network derived from (i) microarray expression profiles, and (ii) prior knowledge about biological mechanisms. PGnet is designed to identify sets of mechanism-anchored genes that are both consistently co-expressed across arrays and differentially expressed between phenotypic conditions.
Arrays and phenotypes
We selected a large array dataset published by the Downing research group (ALL arrays)  that comprises well characterized subtypes of ALL and other clinical phenotypes, including cytogenetic characteristics, molecular status and patient outcomes. The details for leukemia phenotypes and sample size are provided in Suppl. Methods (Additional file 1).
Epigenetic Seed Genes – ESGs (Suppl. Methods (Additional file 1) and Suppl. Table 1 (Additional file 2))
Gene Ontology terms and PubMed were used to identify genes with epigenetic effects. Genes were subsequently mapped to Affymetrix probe-sets and curated into eight categories.
The 132 ALL arrays were normalized with the variance stabilization and calibration normalization (vsn) method  using Bioconductor [32, 33] package compdiagTools . We then applied an additional inter quartile range (IQR) filter  to eliminate genes lacking sufficient variation across samples in expression (Suppl. Methods (Additional file 1)).
Building a phenotype – gene network (PGnet)
Step 1a: Vector of g enes co-e xpressed with the m echanism seed genes
All genes that co-expressed with ESGs were sorted in a vector based on the Pearson correlation coefficient (PCC). Results were denoted as .
Step 1b: Vector of differentially expressed genes in ALL
At the same time, all genes were also sorted based on their adjusted Student t-test conducted between the phenotype of interest against the remaining pooled phenotypes. Results were denoted as . Bioconductor [32, 33] package stats was used to calculate the PCC and the package Twilight  was used to calculate the adjusted t-score parameter.
To compare two ordered lists of gene expressions, we used OrderedList [37–39] from Bioconductor [32, 33], a non-parametric quantitative vectorial enrichment method that we previously published, and has been shown more sensitive to detect significant departure from a predicted distribution than semi-quantitative enrichment approaches such as the Fisher's Exact Test or the the Chi-square test. We calculated a matrix of similarity scores M s = (si, j), where each score si, jassessed the pair-wise similarities between two vectors. The fist vector is the ordered co-expression coefficients V ESGi and the second one is the ordered differential expression statistics VLP j. The similarity score gives higher weights to ranking extremes: the top and bottom ranks in both lists. In our method, we compared the ranking of genes in the co-expression set with those gene ranks from the phenotypic set. This resulted in a total of two comparisons for each phenotype/"seed gene" combination (correlation and anti-correlation).
Step 3: Vectorial Enrichment Optimization (VEO)
For each significant seed gene/phenotype pair we considered these to be "linked." By aggregating these seed gene/phenotype pairs, we developed a tripartite network PGnet.
Step 5: Visualization of the Tripartite Network
Meaning of shapes and colors in the network: triangle (epigenetic seed genes), circle (predicted GEMs) and box (phenotypes); red (up-regulated), blue (down-regulated) and grey for vertex of a gene had more than one linkage and was up-regulated in one condition but down-regulated in a different condition (Suppl. Methods (Additional file 1)). By color-coding the edges of the graph, we are providing a direction to each similarity vector, magenta line for correlation whereas turquoise for anti-correlation. With these vectors, one can judge how these genes express in a specific condition.
Step 6: Biological meaning of the network
Using Gene ontology, we conducted an enrichment of the molecular function and biological processes among the genes identified in the PGnet biomodules in order to characterize biologically the network and we also reviewed the literature for the genes involved in the biomodules associated to BCR-ABL, T-ALL and hyperdiploidy. Thus the resulting set of genes termed as G enes significantly E xpressed with the M echanism (GEMs) in the epigenetic context of this network meet all of the following criteria: (i) they are co-expressed with genes known to be involved in the biological mechanism of interest, (ii) they are also differentially expressed between distinct phenotypes relevant to the study, and (iii) as a biomodule, genes correlate with both the mechanism and the phenotype.
Robust predictive network and evaluation
To demonstrate the accuracy of the derived network to predict relapse, we performed a conservative evaluation consisting of one hundred three-fold cross-validation (CV) studies of the PGnet method. In other words, as shown in Figure 2c, the network was derived from 2/3 of the randomly selected patients and the evaluation was conducted on the remaining third. The random selection was conducted to conserve the respective group sizes (normal, cancer) and was considered a more adequate and severe control . This procedure was repeated one hundred times on different random resamplings. Two different predictive methods were used as well: (i) Prediction Analysis for Microarrays Class Prediction  (PAM) that does not involve any machine learning and (ii) the Support Vector Machine  (SVM) (Suppl. Methods (Additional file 1) and Suppl. Fig. 2 (Additional file 5)). The resulting receiver operating characteristic (ROC) curve, area under the curve (AUC) and corresponding p-values were calculated by the Bioconductor [32, 33] package verification . A robust molecular signature is one that repeatedly appears by random sampling . We further identified the GEMs correlated with the ESGs that were identified as robust . The robust ESGs refer to those GEMs that were among the top 5% frequencies in the one hundred iterations of the 3-fold cross-validation (Figure B in Suppl. Methods (Additional file 1)). Figure 3 illustrates the sub-network associated to the comparison between "Relapse" and the "continuous complete remission – CCR" phenotypes.
Finally we compared our results to those obtained by a straightforward reverse engineering method (ARACNE).
Seventy-one distinct epigenetic seed genes were identified in the literature review and denoted as "seed gene" candidates for input in PGnet (Suppl. Table 1 (Additional file 2)). In the 132 ALL arrays, 7,256 out of 12,997 unique genes and 48 out of the 71 subset of ESGs satisfy the IQR filter (bolded genes in Suppl. Table 1 (Additional file 2)). The 48 ESGs are thus enriched ab initio among genes differentially expressed in ALL samples suggesting a biological relevance in ALL (p = 5%, Fisher's Exact Test).
We built a gene-phenotype network specific for epigenetic genes clusters in ALL as shown in Figures 2a and 2b. The derived network comprises 33 significant nodes, including eight clinical subtypes of ALL and two outcome conditions (LP, vertex in yellow box), 23 epigenetic seed genes (ESG, yellow triangle) and 535 genes that co-express with ESGs (GEM, grey circle). Three of these ESGs and 299 GEMs are up-regulated in association to their phenotype(s) as compared with their expression associated to the remaining pooled phenotypes, while 14 ESGs and 203 GEMs are down-regulated. In addition, 6 ESGs and 33 GEMs were up-regulated in one phenotype but down-regulated in a different phenotype, which are phenotype specifically differential expressions (see Step 3 of Methods). A summary of predictions is provided in Suppl. Table 2 (Additional file 3). These 23 ESGs genes are highly co-expressed with epigenetic genes and also highly differentially expressed in distinct ALL phenotypes groups (CBX1, CBX5, CBX6, CBX7, PHLDA2, BAZ2A, BAZ2B, MYST2, MYST4, MECP2, SMARCA2, HDAC4, HDAC5, HDAC6, HDAC7A, HDAC9, SMARCA4, SMYD3, SUV39H1, DNMT3A, DNMT3B, PRDM2 and MBD2).
To validate the prognostic ability of the genes in PGnet associated with leukemia relapse, we performed one hundred three-fold cross-validations in two ways (Suppl. Methods (Additional file 1) and Suppl. Fig. 2 (Additional file 5)). Using the PAM classification that does not require machine-learning, the predictions were accurate (AUC = 0.65, p = 3%, Suppl. Fig. 3 (Additional file 6)). We also conducted a severe control by randomly selecting genes differentially expressed in the array and the p-values of the derived predictors ranged from 12% to 67%, further corroborating that the epigenetic network derived by PGnet is associated to the relapse outcome. Using SVM machine-learning to improve the predictions in a 3-fold cross over design, PGnet achieved a AUC = 0.67 (p = 1.6%) (Figure 3 in this manuscript). Precision and recall of the predictor in cross-validation studies are also significant (Suppl. Fig. 4 (Additional file 7)). The evaluation confirmed the detection of co-expression biomodules associated to epigenetic alterations can be utilized in the identification of ALL with poor prognosis . Supplementary Table 5 (Additional file 8) reports 52 robust GEMs together with 3 robust ESGs (CBX5, SMARCA4 and DNMT3A) associated with leukemia relapse (Suppl. Methods (Additional file 1)).
We further proceeded to identify biological enrichment in the distinct sets of genes associated with response to therapy and long-term maintenance of disease remission. There were 4 ESGs and a total of 39 GEMs associated with the phenotype "Relapse", such as CCNA2, BUB1, MAD2L1, CDC45L and CCNB2, etc, which were significantly enriched in one GO term: ATP binding (hypergeometric p = 3.5 × 10-6, Suppl. Table 4 (Additional file 9)). Interesting, ATP has been reported as treatment target of murine leukemic cells in vitro to reduce the number of leukemic clonogenic cells .
The GEMs derived from PGnet can distinguish their associated phenotype in addition to be co-expressed with their associated seed gene on transcript level. As an example in this study, Supplementary Figure 5 (Additional file 10) shows that the expression of 61 GEMs can clearly distinguish samples of "Hyperdiploid>50" from other ALL samples. There were 4 ESGs associated with the hyperdiploid karyotypes by PGnet. HDAC6 is a class IIB histon deacetylase and identified as target of anti-leukemia therapy . Inhibition of HDAC6 disrupts the association of HSP90 with its chaperon proteins, resulting in ubiquitylation of certain oncogenes, such as Bcr-Ab l . Three other genes were also down-regulated (SMARCA4, BAZ2A and SMARCC2): SMARCA4 is a drug target candidates in hyperdiploid multiple myeloma ; BAZ2A is a novel nucleolar chromatin remodeling machine , and SMARCC2 was among the top discriminating genes in the good prognosis subgroup of MLL . Moreover, GEMs identified by PGnet significantly enriched among the top-100 marker genes in previous genome-scale studies of ALL (Fisher's test p < 2 × 10-16, Suppl. Table 6 (Additional file 11)).
Comparison of the derived network with other computational methods
ARACNE [24, 53] software was used to reverse engineer the transcriptional network in two ways. First, by providing the genes expression data for entire ALL samples as input, we compared our GEMs with genes identified by ARACNE (Result is given in Suppl. Table 7 (Additional file 12)). We provided ARACNE our 48 epigenetic seed genes and the expression data for samples in each phenotype as input, irrespectively. Subsequently, we got 12 different phenotype specific gene-gene networks. Each ARACNE network detects thousands of genes that with significant (p < 0.05) mutual information (MI) with inputted ESGs. By design, a majority of the genes detected by ARACNE are not phenotype specific. GEMs predicted by PGnet overlap with ARACNE's prediction for 31 of 33 ESGs (Suppl. Table 8 (Additional file 13)). However, PGnet differs from ARACNE in that it provides phenotypic information left out in ARACNE.
ARACNE, FunNet and PGnet provide co-regulation networks as an output and are thus "related"; however, they differ in several important ways: (i) PGnet and FunNet combine supervised technology and non-parametric methodology while ARACNE uses information theory; (ii) inputs to PGnet are expression levels and phenotypic associations of interest such as seed genes whereas FunNet requires full expression together with a reference list of all transcripts to be analyzed and ARACNE uses expression data exclusively; (iii) FunNet abstracts transcriptional functions from co-expression layer; and consequently (iv) PGnet's output is a tripartite network consisting of co-regulated genes and clinical/genetic characteristics of interest while ARACNE's or FunNet's outputs are uni-partite graphs. (v) The significant threshold of PGnet relates to the complete ordering of all genes to be analyzed whereas the significant threshold of FunNet is related to the co-expression of single gene. (vi) PGnet not only provides a degree of association between phenotypes but also sheds light on whether there was concordance in the directionality of the changes in expression level.
PGnet parallelizes two input vectors and finds sets of GEMs via vectorial enrichment optimization. Using measurements of differential expression and co-expression together, PGnet is more reliable in discovering phenotype-specific biomodules that are consistent across every patient than a simplified method that analyzes the expressed pattern of the epigenetic seed genes (ESGs) alone. First, a simplified method identifies none of GEMs from PGnet. However, we have shown some evidence indicating that the GEMs are more likely to be involved in specific epigenetic events than those directly calculated to be correlated to a phenotype of interest. Second, simpler alternate methods relying on co-expression or differential expression separately would identify only a the subset of ESGs from PGnet (data not shown), because these methods use an arbitrary threshold for significance of each gene and neglect the joint analysis of co-expression patterns with those of differential expression. In contrast, the "ESG-phenotype" linkage, which we proposed in PGnet, would be significant even if the epigenetic seed gene itself is not "significantly" differentially expressed in the linked phenotype (for instance, the seed genes that are not self-linked in Suppl. Fig. 1 (Additional file 14)).
Biologically, PGnet is an attractive technique as we know that mechanism-related genes have similar patterns of expression [4, 20], and pathological mechanisms are easier to understand than genes by clinicians. Additionally, the non-parametric rank-correlation algorithm that we previously developed for Bioconductor can use the full range of the expression data for discovery instead of arbitrary statistical cut-offs [37–39]. We have extended it to derive phenotype-genotype correlations based on prior knowledge in addition to gene expression. Moreover, this tri-partite network allows to view genes for which the expression is specific to a phenotype of interest and also anchored to a biological mechanism.
Future studies and limitations
Epigenetic gene regulation is one among many possible mechanisms involved in disease-specific gene aberrant activation. Better predictors of outcome can be developed using a more comprehensive number of biological mechanisms. The PGnet method could be expanded to a broader variety of biological mechanisms in order to provide more accurate mechanism-anchored profiles that predict therapeutic outcome (e.g. transcriptional and microRNA networks , Gene Ontology terms, KEGG, etc), however additional methods are required to control for multiplicity of mechanism while preserving accuracy of the derived tripartite networks. In addition, this PGnet is a supervised method that relies on prior knowledge about seed genes or gene products that regulate epidemic processes. Therefore, PGnet may "skew" the network accordingly, which may reflect only subset of the real regulatory relationship. Further improvement for finding disease associated and seed-gene regulated genes will likely require a refined assessment of co-expression, e.g. mutual information [24, 55], instead of linear Pearson coefficient . By design, PGnet identifies biomodules that are consistently co-expressed with the mechanism seed genes across all patient samples. However, there could exist mechanisms that are only co-expressed in some specific phenotypes and otherwise the co-expression patterns are lost. These particular biomodules may also contribute to mechanism-anchored predictors and require further methodological developments for their ascertainment. Future evaluations comparing the PGnet-derived predictors in other datasets are required, and we intend to proceed with multi-mechanism profiling that would in theory achieve higher precision and recall.
We introduced and evaluated a novel algorithm, PGnet, to identify mechanism-anchored co-expression networks and to predict therapeutic outcome. PGnet differs from previous reverse engineering methods in that it provides a more comprehensive output consisting of a tripartite network of expression similarity between genes, biological mechanisms and clinical phenotypes. Additionally, statistical significance is conducted over expression ordering inclusive of the complete array.
Trained on epigenetic mechanisms, PGnet accurately classified patients in the leukemia subtype and the relapse group, and these results suggest that a more comprehensive multi mechanism-based profile may achieve higher accuracy scores. The proposed method is scalable, in principle, to other mechanisms such as transcriptional networks, microRNA-regulated or Gene Ontology classes. In addition, the produced "similarity linkages" between mechanisms and genes comprise magnitude and direction (correlated or anti-correlated), which could also be utilize to infer regulation (activation or suppression) .
This work was supported in part by the The Cancer Research Foundation, the NIH National Center for Multiscale Analyses of Genomic and Cellular Networks (MAGNET) 1U54CA121852, and the Natural Science Foundation of China 60771024, 60121101 and. We thank Dr. Richard Sheuermann for his contribution to the interpretation of the network. We also thank Dr. Zuhong Lu for his contribution in developing the epigenetic gene list. We would also like to acknowledge Matthew Crowson for his assistance in the revision of this manuscript.
This article has been published as part of BMC Bioinformatics Volume 10 Supplement 9, 2009: Proceedings of the 2009 AMIA Summit on Translational Bioinformatics. The full contents of the supplement are available online at http://www.biomedcentral.com/1471-2105/10?issue=S9.
- Michiels S, Koscielny S, Hill C: Prediction of cancer outcome with microarrays: a multiple random validation strategy. Lancet 2005, 365(9458):488–492. 10.1016/S0140-6736(05)17866-0View ArticlePubMedGoogle Scholar
- Lottaz C, Spang R: Molecular decomposition of complex clinical phenotypes using biologically structured analysis of microarray data. Bioinformatics 2005, 21(9):1971–1978. 10.1093/bioinformatics/bti292View ArticlePubMedGoogle Scholar
- Pan W: Incorporating gene functions as priors in model-based clustering of microarray gene expression data. Bioinformatics 2006, 22(7):795–801. 10.1093/bioinformatics/btl011View ArticlePubMedGoogle Scholar
- Barabasi AL, Oltvai ZN: Network biology: understanding the cell's functional organization. Nat Rev Genet 2004, 5(2):101–113. 10.1038/nrg1272View ArticlePubMedGoogle Scholar
- Pujana MA, Han JD, Starita LM, Stevens KN, Tewari M, Ahn JS, Rennert G, Moreno V, Kirchhoff T, Gold B, et al.: Network modeling links breast cancer susceptibility and centrosome dysfunction. Nat Genet 2007, 39(11):1338–1349. 10.1038/ng.2007.2View ArticlePubMedGoogle Scholar
- Dillon N: Gene regulation and large-scale chromatin organization in the nucleus. Chromosome Res 2006, 14(1):117–126. 10.1007/s10577-006-1027-8View ArticlePubMedGoogle Scholar
- Weber M, Schubeler D: Genomic patterns of DNA methylation: targets and function of an epigenetic mark. Curr Opin Cell Biol 2007, 19(3):273–280. 10.1016/j.ceb.2007.04.011View ArticlePubMedGoogle Scholar
- Gelato KA, Fischle W: Role of histone modifications in defining chromatin structure and function. Biol Chem 2008, 389(4):353–363. 10.1515/BC.2008.048View ArticlePubMedGoogle Scholar
- Gangaraju VK, Bartholomew B: Mechanisms of ATP dependent chromatin remodeling. Mutat Res 2007, 618(1–2):3–17.PubMed CentralView ArticlePubMedGoogle Scholar
- Lopez-Serra L, Esteller M: Proteins that bind methylated DNA and human cancer: reading the wrong words. Br J Cancer 2008, 98(12):1881–1885. 10.1038/sj.bjc.6604374PubMed CentralView ArticlePubMedGoogle Scholar
- Burgess R, Jenkins R, Zhang Z: Epigenetic changes in gliomas. Cancer Biol Ther 2008, 7(9):1326–1334.PubMed CentralView ArticlePubMedGoogle Scholar
- Feinberg AP, Tycko B: The history of cancer epigenetics. Nat Rev Cancer 2004, 4(2):143–153. 10.1038/nrc1279View ArticlePubMedGoogle Scholar
- Gore SD: Combination therapy with DNA methyltransferase inhibitors in hematologic malignancies. Nat Clin Pract Oncol 2005, 2(Suppl 1):S30–35. 10.1038/ncponc0346View ArticlePubMedGoogle Scholar
- Jones PA, Baylin SB: The epigenomics of cancer. Cell 2007, 128(4):683–692. 10.1016/j.cell.2007.01.029PubMed CentralView ArticlePubMedGoogle Scholar
- Bock C, Lengauer T: Computational epigenetics. Bioinformatics 2008, 24(1):1–10. 10.1093/bioinformatics/btm546View ArticlePubMedGoogle Scholar
- Kann MG: Protein interactions and disease: computational approaches to uncover the etiology of diseases. Brief Bioinform 2007, 8(5):333–346. 10.1093/bib/bbm031View ArticlePubMedGoogle Scholar
- Loscalzo J, Kohane I, Barabasi AL: Human disease classification in the postgenomic era: a complex systems approach to human pathobiology. Mol Syst Biol 2007, 3: 124. 10.1038/msb4100163PubMed CentralView ArticlePubMedGoogle Scholar
- van Someren EP, Wessels LF, Backer E, Reinders MJ: Genetic network modeling. Pharmacogenomics 2002, 3(4):507–525. 10.1517/146224188.8.131.527View ArticlePubMedGoogle Scholar
- McKinney BA, Reif DM, Ritchie MD, Moore JH: Machine learning for detecting gene-gene interactions: a review. Appl Bioinformatics 2006, 5(2):77–88. 10.2165/00822942-200605020-00002PubMed CentralView ArticlePubMedGoogle Scholar
- Oti M, Brunner HG: The modular nature of genetic diseases. Clin Genet 2007, 71(1):1–11. 10.1111/j.1399-0004.2006.00708.xView ArticlePubMedGoogle Scholar
- Ideker T, Sharan R: Protein networks in disease. Genome Res 2008, 18(4):644–652. 10.1101/gr.071852.107PubMed CentralView ArticlePubMedGoogle Scholar
- Lussier YA, Liu Y: Computational approaches to phenotyping: high-throughput phenomics. Proc Am Thorac Soc 2007, 4(1):18–25. 10.1513/pats.200607-142JGPubMed CentralView ArticlePubMedGoogle Scholar
- Lewin R: First success with reverse genetics. Science 1986, 233(4760):159–160.Google Scholar
- Margolin AA, Wang K, Lim WK, Kustagi M, Nemenman I, Califano A: Reverse engineering cellular networks. Nat Protoc 2006, 1(2):662–671. 10.1038/nprot.2006.106View ArticlePubMedGoogle Scholar
- Ma S-K: Statistical mechanics. Singapore: World Scientific; 1985.View ArticleGoogle Scholar
- Blake JA, Harris MA: The Gene Ontology (GO) project: structured vocabularies for molecular biology and their application to genome and expression analysis. Curr Protoc Bioinformatics 2002., Chapter 7(Unit 7 2):Google Scholar
- Blake JA, Harris MA: The Gene Ontology (GO) project: structured vocabularies for molecular biology and their application to genome and expression analysis. Curr Protoc Bioinformatics 2008., Chapter 7(Unit 7 2):Google Scholar
- Kanehisa M, Goto S, Kawashima S, Nakaya A: The KEGG databases at GenomeNet. Nucleic Acids Res 2002, (30):42–46. 10.1093/nar/30.1.42Google Scholar
- Prifti E, Zucker JD, Clement K, Henegar C: FunNet: an integrative tool for exploring transcriptional interactions. Bioinformatics 2008, 24(22):2636–2638. 10.1093/bioinformatics/btn492View ArticlePubMedGoogle Scholar
- Ross ME, Zhou X, Song G, Shurtleff SA, Girtman K, Williams WK, Liu HC, Mahfouz R, Raimondi SC, Lenny N, et al.: Classification of pediatric acute lymphoblastic leukemia by gene expression profiling. Blood 2003, 102(8):2951–2959. 10.1182/blood-2003-01-0338View ArticlePubMedGoogle Scholar
- Huber W, von Heydebreck A, Sultmann H, Poustka A, Vingron M: Variance stabilization applied to microarray data calibration and to the quantification of differential expression. Bioinformatics 2002, 18(Suppl 1):S96–104.View ArticlePubMedGoogle Scholar
- Gentleman RC, Carey VJ, Bates DM, Bolstad B, Dettling M, Dudoit S, Ellis B, Gautier L, Ge Y, Gentry J, et al.: Bioconductor: open software development for computational biology and bioinformatics. Genome Biol 2004, 5(10):R80. 10.1186/gb-2004-5-10-r80PubMed CentralView ArticlePubMedGoogle Scholar
- R-Development-Core-Team: R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing,: 2005; Vienna, Austria 2005.Google Scholar
- Scheid S, Jäger J, Lottaz C: A toolbox for performing and illustrating microarray data analyses – user's guide to the R package compdiagTools. CompDiag Tech Rep 2005.Google Scholar
- von Heydebreck A, Huber W, Gentleman R: Differential expression with the Bioconductor Project. Bioconductor Project Working Papers 2004.Google Scholar
- Scheid S, Spang R: twilight; a Bioconductor package for estimating the local false discovery rate. Bioinformatics 2005, 21(12):2921–2922. 10.1093/bioinformatics/bti436View ArticlePubMedGoogle Scholar
- Yang X, Bentink S, Scheid S, Spang R: Similarities of ordered gene lists. J Bioinform Comput Biol 2006, 4(3):693–708. 10.1142/S0219720006002120View ArticlePubMedGoogle Scholar
- Yang X, Sun X: Meta-analysis of several gene lists for distinct types of cancer: a simple way to reveal common prognostic markers. BMC Bioinformatics 2007, 8: 118. 10.1186/1471-2105-8-118PubMed CentralView ArticlePubMedGoogle Scholar
- Lottaz C, Yang X, Scheid S, Spang R: OrderedList – a bioconductor package for detecting similarity in ordered gene lists. Bioinformatics 2006, 22(18):2315–2316. 10.1093/bioinformatics/btl385View ArticlePubMedGoogle Scholar
- Storey J: A direct approach to false discovery rates. Journal of the Royal Statistical Society, Series B 2002, 64: 479–498. 10.1111/1467-9868.00346View ArticleGoogle Scholar
- Storey J: The positive false discovery rate: A Bayesian interpretation and the q-value. Annals of Statistics 2003, 31: 2013–2035. 10.1214/aos/1074290335View ArticleGoogle Scholar
- Barnett V: Sample Survey Principles and Method (The second edition). Hodder Arnold; 1991.Google Scholar
- Tibshirani R, Hastie T, Narasimhan B, Chu G: Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc Natl Acad Sci USA 2002, 99(10):6567–6572. 10.1073/pnas.082099299PubMed CentralView ArticlePubMedGoogle Scholar
- Meye D, Leisch F, Hornik K: The support vector machine under test. Neurocomputing 2003, 55(1–2):169–186.Google Scholar
- Mason SJ, Graham NE: Areas beneath the relative operating characteristics (ROC) and relative operating levels (ROL) curves: Statistical significance and interpretation. Q J R Meteorol Soc 2002, 128: 2145–2166. 10.1256/003590002320603584View ArticleGoogle Scholar
- Yang H, Kadia T, Xiao L, Bueso-Ramos CE, Hoshino K, Thomas DA, O'Brien S, Jabbour E, Pierce S, Rosner GL, et al.: Residual DNA methylation at remission is prognostic in adult Philadelphia chromosome-negative acute lymphocytic leukemia. Blood 2009, 113(9):1892–1898. 10.1182/blood-2008-02-141002PubMed CentralView ArticlePubMedGoogle Scholar
- Hatta Y, Itoh T, Baba M, Miyajima T, Shimojima H, Sawada U, Horie T: Purging in autologous hematopoietic stem cell transplantation using adenosine triphosphate (ATP) and 4-hydroperoxycyclophosphamide (4-HC). Leuk Res 2002, 26(5):477–482. 10.1016/S0145-2126(01)00164-3View ArticlePubMedGoogle Scholar
- Rao R, Fiskus W, Yang Y, Lee P, Joshi R, Fernandez P, Mandawat A, Atadja P, Bradner JE, Bhalla K: HDAC6 inhibition enhances 17-AAG – mediated abrogation of hsp90 chaperone function in human leukemia cells. Blood 2008, 112(5):1886–1893. 10.1182/blood-2008-03-143644View ArticlePubMedGoogle Scholar
- Bali P, Pranpat M, Bradner J, Balasis M, Fiskus W, Guo F, Rocha K, Kumaraswamy S, Boyapalle S, Atadja P, et al.: Inhibition of histone deacetylase 6 acetylates and disrupts the chaperone function of heat shock protein 90: a novel basis for antileukemia activity of histone deacetylase inhibitors. J Biol Chem 2005, 280(29):26729–26734. 10.1074/jbc.C500186200View ArticlePubMedGoogle Scholar
- Carrasco DR, Tonon G, Huang Y, Zhang Y, Sinha R, Feng B, Stewart JP, Zhan F, Khatry D, Protopopova M, et al.: High-resolution genomic profiles define distinct clinico-pathogenetic subgroups of multiple myeloma patients. Cancer Cell 2006, 9(4):313–325. 10.1016/j.ccr.2006.03.019View ArticlePubMedGoogle Scholar
- Strohner R, Nemeth A, Jansa P, Hofmann-Rohrer U, Santoro R, Langst G, Grummt I: NoRC – a novel member of mammalian ISWI-containing chromatin remodeling machines. Embo J 2001, 20(17):4892–4900. 10.1093/emboj/20.17.4892PubMed CentralView ArticlePubMedGoogle Scholar
- Tsutsumi S, Taketani T, Nishimura K, Ge X, Taki T, Sugita K, Ishii E, Hanada R, Ohki M, Aburatani H, et al.: Two distinct gene expression signatures in pediatric acute lymphoblastic leukemia with MLL rearrangements. Cancer Res 2003, 63(16):4882–4887.PubMedGoogle Scholar
- Basso K, Margolin AA, Stolovitzky G, Klein U, Dalla-Favera R, Califano A: Reverse engineering of regulatory networks in human B cells. Nat Genet 2005, 37(4):382–390. 10.1038/ng1532View ArticlePubMedGoogle Scholar
- Gennarino VA, Sardiello M, Avellino R, Meola N, Maselli V, Anand S, Cutillo L, Ballabio A, Banfi S: MicroRNA target prediction by expression analysis of host genes. Genome Res 2008.Google Scholar
- Margolin AA, Nemenman I, Basso K, Wiggins C, Stolovitzky G, Dalla Favera R, Califano A: ARACNE: an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context. BMC Bioinformatics 2006, 7(Suppl 1):S7. 10.1186/1471-2105-7-S1-S7PubMed CentralView ArticlePubMedGoogle 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.