Previous Article |
Table of Contents
| Next Article 
BIOLOGICAL SCIENCES / MEDICAL SCIENCES
Analysis of oncogenic signaling networks in glioblastoma identifies ASPM as a molecular target
Departments of dPathology and Laboratory Medicine, lNeurology, hNeurosurgery, aHuman Genetics, fPharmacology, and bBiostatistics, iThe Henry E. Singleton Brain Cancer Research Program and jNeurogenetics Research Program, and the kSemel Institute for Neuroscience at the David Geffen School of Medicine, University of California, Los Angeles, CA 90095; eIngenuity Systems, Inc., 1700 Seaport Boulevard, Third Floor, Redwood City, CA 94063; gThe Barrows Neurological Institute, St. Joseph's Hospital–Catholic Healthcare West, 350 West Thomas Road, Phoenix, AZ 85013; and mDepartments of Medicine and Molecular Genetics and Microbiology, Institute for Genome Sciences and Policy, 101 Science Drive, Duke University Medical Center, Durham, NC 27708
Communicated by Michael E. Phelps, University of California School of Medicine, Los Angeles, CA, September 29, 2006 (received for review June 22, 2006)
 |
Abstract |
|---|
 Top Abstract ResultsDiscussionMethodsAcknowledgementsReferences |
|---|
epidermal growth factor receptor vIII | Glioblastoma | modules | weighted gene coexpression network analysis | network-based screening
The wealth of molecular information provided by genomic technologies provides a remarkable opportunity for new target discovery (3, 4). Gene expression data can provide a key first step toward constructing a systems level view of the perturbed networks in cancer cells, thus potentially identifying key genes, networks, or pathways that can be therapeutically targeted (5). However, the identification of key molecular targets still remains a challenge. Recent work highlights the potential for uncovering oncogenic pathways and molecular targets, when genomic data are analyzed at the level of gene coexpression modules or metagenes or when aggregated gene sets are used to assess modules enriched for key biological processes (6–8). Integrating this type of data with studies in model systems in which modules can be studied in response to relevant molecular perturbations (e.g., oncogene overexpression or pharmacological inhibition) may further facilitate the identification and validation of novel molecular targets (9–11). Here, we adopt an unbiased strategy to detect an oncogenic module in glioblastoma and integrate this with studies in isogenic cell systems to identify and validate ASPM (abnormal spindle-like microcephaly associated) as a previously undescribed glioblastoma target.
|
Results |
|---|
|
TopAbstractResultsDiscussionMethodsAcknowledgementsReferences |
|---|
Five gene coexpression modules were detected in glioblastoma data set 1 (Fig. 1a). These modules were significantly enriched for genes with the following specific ontologic classes: (i) mitosis/cell cycle (185 genes, P = 7.2 x 10–42); (ii) immune response (606 genes, P = 2.4 x 10–36); (iii) neurogenesis (143 genes, P = 4.0 x 10–4); (iv) cytoplasm (1,112 genes P = 1.1 x 10–12); and (v) metabolism (136 genes, P = 1.8 x 10–2) (EASE software: http://david.niaid.nih.gov/david/ease1.htm) (Fig. 1d). The fact that unsupervised clustering based on a coexpression measure resulted in modules enriched for biologically important processes, including cancer-related themes, suggests that these modules are a robust feature of the molecular architecture of glioblastoma.
|
2 test, P = 2.2 x 10–16, Rand index measure of agreement = 0.9) (Fig. 1b). Thus, gene coexpression modules are highly preserved in both glioblastoma data sets. To determine whether these modules were detectable in another cancer type, we analyzed a publicly available breast cancer data set (14). This data set was sufficiently large and contained gene expression data from a different microarray platform, allowing for array platform independent conclusions. Probe sets that were common to both array platforms were mapped, and a weighted gene coexpression network was constructed based for the breast cancer data set. To determine whether glioblastoma modules were present in breast cancer, we assigned glioblastoma module colors to the genes in the hierarchical clustering tree of the breast cancer data. This revealed that two glioblastoma modules (the cell cycle/mitosis module and the immune response module) were highly preserved (Fig. 1c) (data available upon request). This suggested that the cell cycle/mitosis module may be involved in biological processes that are shared by both cancer types.
To further assess whether the mitosis/cell cycle module (MCM) is present in other cancers, we determined whether the genes contained within the module are part of the "metasignatures" of cancer that have been derived from metaanalyses of a large numbers of samples of different cancer types (15). Two metasignatures (MS) of cancer have been identified by large-scale metaanalyses across multiple cancer types, including a MS of "undifferentiated cancer" (15). Of the 48 genes of the "undifferentiated cancer" MS that were present in the glioblastoma data set, 33 (69%) were present in the MCM (P = 2.7 x 10–31).
To correlate individual expression profiles with the entire module, we summarized the expression profile of the module genes by the first module eigengene, which is defined by using the singular value decomposition of the expression data (16). To determine whether this MCM is a proliferation cluster, we correlated the module eigengene with Ki67 and PCNA (two clinically used markers of cancer cell proliferation and members of the module) (17). The module eigengene was highly correlated with both Ki67 and PCNA (Ki67: r = 0.74; P < 6.2 x 10–7 for data set 1; and r = 0.81; P < 1 x 10–20 for data set 2; PCNA: r = 0.79 P < 1 x 10–20 for data set 1; and r = 0.80; P < 1 x 10–20 for data set 2) (Fig. 5 a–d, which is published as supporting information on the PNAS web site). We next examined the expression pattern of these module genes across 353 samples including glioblastoma and other tissues (both tumor and normal). Expression of MCM genes goes up or down together across a wide range of tissue types, including glioblastoma, meningioma, normal brain, fetal brain, a range of normal nonbrain tissues, and a range of fetal nonbrain tissues (Fig. 1e). The pattern of expression of this module in a subset of glioblastomas is quite similar to that of fetal tissues (including fetal brain), and quite unlike that of normal mature brain or body tissues. Further, this module is highly expressed in only a subset of glioblastomas, the type 2A pattern, which we have shown to be associated with poor prognosis (12). The module is not highly expressed in some other subsets of glioblastoma, including the type 2B pattern that we have shown to be a highly aggressive tumor type as well (12). This raises the possibility that this proliferation module is specific to a subset of glioblastomas and that it shares similarity with a fetal proliferation signature.
Highly connected "hub" genes are thought to play an important role in organizing the behavior of biological modules (18–21). Therefore, we set out to identify the MCM hubs (22). We defined a connectivity measure (K) for each gene based on its Pearson correlation with all of the other genes in the module as described in Methods (13). Because highly connected hub genes are far more likely than nonhub genes to be essential for survival in lower organisms (18–20), we hypothesized that intramodular hub genes may be associated with survival in cancer. To define a measure of prognostic significance, we used a univariate Cox proportional hazards regression model to regress patient survival on the individual gene expression profiles. The resulting univariate Cox-regression p-values were used to define a measure of prognostic significance as follows: GS = –log10(Cox P value), i.e., this measure of gene significance is proportional to the number of zeroes in the P value. In the MCM, intramodular connectivity K and prognostic significance GS, were significantly correlated in both glioblastoma data sets (r = 0.59, P = 7.1 x 10–19 in data set 1, and r = 0.59, P = 6.5 x 10–19 in data set 2) (Fig. 2a and b).
|
Some of the most highly connected genes within the MCM already have been identified as potential cancer targets (topoisomerase II
, ARKB, PTTG1/sercuin, Survivin, and EZH2) (23–27). To identify a potential novel gene target, we looked for the most highly connected genes that have not been extensively studied as cancer targets. This led us to study the ASPM gene, because it had the highest K value in both glioblastoma data sets of any gene that has not been previously recognized as a cancer target. ASPM is the human ortholog of a Drosophila mitotic spindle protein, encoding the protein microcephalin (28–30). ASPM is thought to regulate neuroblast proliferation (29), and it has recently been shown to be a key regulator of brain size through evolution (31–33). Mutations within this gene are associated with primary human microcephaly (29, 30). A recent study demonstrated increased ASPM in ovarian and uterine cancers, suggesting that it may play a role in other cancer types (34), although it was not part of the MS of undifferentiated cancer (15), suggesting that it may have specificity for only a few tumor types including glioblastoma. Because ASPM is expressed at a very low level in normal brain (and normal body tissues) relative to glioblastoma (Fig. 1f), we reasoned that it could present a compelling molecular target.
Strikingly, the traditional proliferation markers Ki67 (Cox regression P = 0.13) and PCNA (P = 0.021) were less associated with glioblastoma survival than 9 of the top 10 most connected hub genes, including ASPM. Specifically, for the combined glioblastoma data set, the P values in the univariate Cox model for these nine hub genes were as follows: TOP2A (P = 0.00088), RACGAP1 (P = 0.0022), KIF4A (P = 0.0030), TPX2 (P = 0.0021), CDC2 (P = 0.0072), EZH2 (P = 0.024), CDC20 (P = 0.0029), KIF14 (P = 0.0020), RAMP (P = 0.015), and ASPM (P = 0.0059). These results suggest that the hub genes, including ASPM, may be more predictive of clinical outcome than the traditional markers of proliferation, PCNA and Ki67. Further, we found that the top third most connected genes encode proteins that are known to interact physically and/or functionally to regulate metaphase to anaphase transition (Fig. 6, which is published as supporting information on the PNAS web site).
To identify a potential molecular mechanism underlying regulation of the MCM, we used a series of isogenic U87 glioblastoma cells engineered to express EGFR, EGFRvIII, and PTEN in relevant combinations (2) (Fig. 3). These molecular alterations are common in glioblastoma (35–37), and they play a critical role in determining response to EGFR kinase inhibitor therapy (2). We performed global transcriptional profiling from RNA that was extracted from duplicate cultures of each of the isogenic U87 lines and analyzed the expression of the MCM. MCM genes were significantly up-regulated in the EGFRvIII overexpressing cells (Fig. 3a). Of note, these are also the most proliferative of the cells. These data suggested that the module is potentially downstream of EGFRvIII signaling; possibly via the PI3K pathway signaling because PTEN coexpression inhibited up-regulation of this module. We therefore analyzed expression of a series of these genes in response to the EGFR inhibitor erlotinib. Expression of each of six representative hub genes tested (ASPM, PRC1, ARKB, MELK, PTTG1, and TOP2) was increased by EGFRvIII, which was abrogated by treatment with the EGFR inhibitor erlotinib (Fig. 3b). Thus, up-regulation of the MCM is regulated by EGFRvIII signaling in glioblastoma cells, likely via its ability to confer a proliferative advantage to these cells.
|
|
|
Discussion |
|---|
|
TopAbstractResultsDiscussionMethodsAcknowledgementsReferences |
|---|
WGCNA also alleviates the multiple testing problem inherent in microarray data analysis. Instead of relating thousands of genes to the clinical outcome, it focuses on the relationship between a few (here 5) modules and the clinical trait. It is worth repeating that the modules were constructed in an unsupervised manner, i.e., without regard to the clinical outcome. Because the modules may correspond to biological pathways, focusing the analysis on module eigengenes (and equivalently intramodular hub genes) amounts to a biologically motivated data reduction scheme. WGCNA starts from the level of thousands of genes, identifies clinically interesting gene modules, and finally uses intramodular connectivity to suggest suitable targets. Because the expression profiles of intramodular hub genes inside an interesting module are highly correlated (in our data, r > 0.90) typically dozens of targets result. Although these targets are statistically equivalent, they may differ in terms of biological plausibility or clinical utility. In many applications, the list of module hub genes may be further winnowed down based on (i) biological plausibility based on external gene (ontology) information, (ii) the availability of protein biomarkers for further validation, (iii) the availability of suitable mouse models for further validation, and/or, (iv) the druggability, i.e., the opportunity for therapeutic intervention.
Understanding how broad cancer-related modules interact with specific molecular lesions in an individual patient is a critical step in finding new molecular targets. Our finding that the MCM is downstream of EGFRvIII signaling suggests a potentially important link by which this process is switched on by an upstream molecular lesion. It is not surprising that the EGFRvIII expressing, PTEN-deficient glioblastoma cells also were the most proliferative. Many of the hub genes identified here (ASPM, BUB1, HEK, STK6, NEK2, PTTG1, PRC1, KNSL2, CCNB1, CDC2, and CDC20) also have been shown to be downstream of other key molecular lesions such as BRCA1 in breast cancer (42) or the human papilloma virus proteins E6/E7 in patients with cervival cancer (for which a "proliferation cluster" of 123 genes associated with HPV E6/E7 in clinical samples strikingly overlapped with the glioblastoma mitosis/cell cycle module; P = 2.2 x 10–16) (43). In addition, there was a highly significant overlap with genes that have been shown to be highly overexpressed in high grade breast cancer (P = 3.6 x 10–75) (44). These data indicate that the MCM may be up-regulated by a number of key molecular lesions that confer a proliferation advantage, thus raising the possibility that common therapies targeting this module may be useful in patients with different types of aggressive cancer.
ASPM had the highest connectivity index in both glioblastoma data sets for any gene not already known to be a cancer target, and it is expressed at very low level in normal brain (and normal body tissues) relative to glioblastoma. ASPM also has been recognized as a critical regulator of brain size, likely via its role in promoting neuroblast proliferation and symmetric division (28–30, 45). Our data showing that neural stem cell differentiation results in loss of ASPM expression and that siRNA-mediated knockdown of ASPM specifically inhibits neural stem cell self renewal and glioblastoma growth suggests the possibility that this gene may be involved in glioblastoma pathogenesis by promoting a stem cell phenotype. Further studies will be necessary to examine the suitability of targeting ASPM in glioblastomas and to determine whether it mediates its effects on glioblastoma by promoting cancer stem cell self-renewal. In summary, this study provides a blueprint for using genomic data to identify key control networks and molecular targets for glioblastoma and, potentially, for other cancers.
|
Methods |
|---|
|
TopAbstractResultsDiscussionMethodsAcknowledgementsReferences |
|---|
Genomic and Functional Analysis in Glioblastoma Cell Lines. The isogenic U87MG expressing PTEN, EGFR, and EGFRvIII in varying combinations have been reported in ref. 2. In brief, cell lines were grown in duplicate cultures under serum free conditions for 48 h, and RNA was isolated by using the Qiagen (Valencia, CA) RNeasy Mini Kit Gene. Expression analysis by using Affymetrix HG-U133A arrays was performed and analyzed, as described above.
EGFR Inhibitor Treatment and siRNA Studies. The EGFR tyrosine kinase inhibitor Erlotinib (Tarceva, OSI-774) was kindly provided by Genentech (South San Francisco, CA). U87MG and U87-EGFRvIII cells (1 x 105) were seeded, respectively, in 100-mm culture dishes and maintained in DMEM supplemented with 10% FBS. Cells were incubated in 5% CO2, 95% humidity incubator for 3 days to reach 50–70% confluency. Then all cells were switched to serum-free medium. The next day U87-EGFRvIII cells were treated by 5 µM OSI-774, whereas U87MG and U87-EGFRvIII control group received the equivalent vehicle. Twenty-four hours later, cell total RNA was isolated by Qiagen RNeasy Mini Kit. RT-PCR analysis of expression of selected genes after treatment is described in the Supporting Methods, which is published as supporting information on the PNAS web site. The specific methods for siRNA studies are available in Supporting Methods. For proliferation assays, 1,500 cells per well in eight replicates were seeded into 96-well plates. Cells were fixed and stained by 0.25% crystal violet in methanol every day or every other day. Stained plates were densitometry scored by AlphaImager 2200 software and plot in Microsoft (Redmond, WA) Excel.
Neurosphere Cell Culture and Transfection. Cerebral cortex was isolated from embryonic day 12 mice. Cells were dissociated and cultured at 50,000 cells/ml in neurosphere formation medium [Neural Basal medium (Invitrogen, Carlsbad, CA) with B27 (GIBCO BRL, Carlsbad, CA), basic fibroblast growth factor (Peprotech, Rocky Hill, NJ), EGF (Chemicon, Temecula, CA), heparin (Sigma-Aldrich, St. Louis, MO) and penicillin-streptomycin (Gemini Bioproducts, West Sacramento, CA)] for a week. Growth factors were added every 3 days. Neurospheres were dissociated and plated onto polyL-ornithine (Sigma)/fibronectin-coated six-well plates in neural basal medium with 2% FBS (GIBCO BRL). Six hours later, the serum medium was removed and replaced with neurosphere formation medium without heparin and penicillin-streptomycin. Twenty-four hours later, cells were transfected with 100 nM siRNA targeting ASPM and 100 nM control siRNA targeting firefly luciferase by using lipofectAMINE 2000 (Invitrogen). The cells were incubated with reagents for 6 h and passaged for secondary neurosphere formation assay.
Secondary Neurosphere Formation Assay. Cells were lifted off the plate with TriplExpress (GIBCO BRL) and then placed into neurosphere formation medium at 1,000 cells/ml and 100 cells/ml. Neurospheres were propagated for 1 week, and the number and the size of the secondary neurospheres formed were measured by using Microcomputer Imaging Device program.
A comprehensive materials and methods section is available upon request. We also provide the entire statistical code, the data, and a weighted gene coexpression network analysis tutorial so that the reader can reproduce all of our findings.
|
Acknowledgements |
|---|
|
TopAbstractResultsDiscussionMethodsAcknowledgementsReferences |
|---|
|
Footnotes |
|---|
Abbreviations: EGFR, epidermal growth factor receptor; MCM, mitosis/cell cycle module; MS, metasignatures.
cTo whom correspondence should be addressed. E-mail: pmischel{at}mednet.ucla.edu or snelson{at}ucla.edu Correspondence regarding statistical issues should be addressed to S.H. E-mail: shorvath{at}mednet.ucla.edu
Freely available online through the PNAS open access option.
Author contributions: S.H., M.C., H.I.K., T.F.C., S.F.N., and P.S.M. designed research; S.H., B.Z., M.C., K.V.L., S.Z., W.Z., S.Q., Z.C., Y.L., A.C.S., L.M.L., P.G.F., H.I.K., T.F.C., S.F.N., and P.S.M. performed research; K.V.L., R.M.F., M.F.L., A.C.S., L.M.L., H.I.K., T.F.C., S.F.N., and P.S.M. contributed new reagents/analytic tools; S.H., Z.B., S.Z., R.M.F., W.Z., S.Q., Z.C., Y.L., H.W., D.H.G., P.G.F., H.I.K., T.F.C., S.F.N., and P.S.M. analyzed data; and S.H., M.C., H.W., D.H.G., P.G.F., H.I.K., T.F.C., S.F.N., and P.S.M. wrote the paper.
The authors declare no conflict of interest.
© 2006 by The National Academy of Sciences of the USA
|
References |
|---|
|
TopAbstractResultsDiscussionMethodsAcknowledgementsReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg What's this?
This article has been cited by other articles in HighWire Press-hosted journals:
|
|
 |
|
  M. P. Keller, Y. Choi, P. Wang, D. Belt Davis, M. E. Rabaglia, A. T. Oler, D. S. Stapleton, C. Argmann, K. L. Schueler, S. Edwards, et al. A gene expression network model of type 2 diabetes links cell cycle regulation in islets with diabetes susceptibility Genome Res., May 1, 2008; 18(5): 706 - 716. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 C. Li and H. Li Network-constrained regularization and variable selection for analysis of genomic data Bioinformatics, May 1, 2008; 24(9): 1175 - 1182. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. A. Miller, M. C. Oldham, and D. H. Geschwind A Systems Level Analysis of Transcriptional Changes in Alzheimer's Disease and Normal Aging J. Neurosci., February 6, 2008; 28(6): 1410 - 1420. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 F. B. Furnari, T. Fenton, R. M. Bachoo, A. Mukasa, J. M. Stommel, A. Stegh, W. C. Hahn, K. L. Ligon, D. N. Louis, C. Brennan, et al. Malignant astrocytic glioma: genetics, biology, and paths to treatment Genes & Dev., November 1, 2007; 21(21): 2683 - 2710. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. Albert Network Inference, Analysis, and Modeling in Systems Biology PLANT CELL, November 1, 2007; 19(11): 3327 - 3338. [Full Text] [PDF]
|
|
|
|
 |
|
 J. K. Lee, D. M. Havaleshko, H. Cho, J. N. Weinstein, E. P. Kaldjian, J. Karpovich, A. Grimshaw, and D. Theodorescu A strategy for predicting the chemosensitivity of human cancers and its application to drug discovery PNAS, August 7, 2007; 104(32): 13086 - 13091. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
