|Home | About | Journals | Submit | Contact Us | Français|
HOXA genes encode critical transcriptional regulators of embryonic development that have been implicated in cancer. In this study, we documented functional relevance and mechanism of activation of HOXA9 in glioblastoma (GBM), the most common malignant brain tumor. Expression of HOXA genes was investigated using RT-PCR in primary gliomas and glioblastoma cell lines and was validated in two sets of expression array data. In a subset of GBM, HOXA genes are aberrantly activated within confined chromosomal domains. Transcriptional activation of the HOXA cluster was reversible by a PI3K inhibitor through an epigenetic mechanism involving histone H3K27 trimethylation. Functional studies of HOXA9 showed its capacity to decrease apoptosis and increase cellular proliferation along with TRAIL resistance. Notably, aberrant expression of HOXA9 was independently predictive of shorter overall and progression-free survival in two GBM patient sets, and improved survival prediction by MGMT promoter methylation. Thus, HOXA9 activation is a novel, independent and negative prognostic marker in GBM that is reversible through a PI3K-associated epigenetic mechanism. Our findings suggest a transcriptional pathway through which PI3K activates oncogenic HOXA expression with implications for mTOR or PI3K targeted therapies.
Glioblastoma (GBM) is the most common and malignant primary brain tumor in adults. With rare exception, the infiltrative nature of GBM precludes surgical cure and adjuvant therapies have achieved only modest benefit, with median survival remaining approximately 12 months (1). Specific molecular characteristics of GBM can be used to define prognostic subgroups (2-4). MGMT promoter methylation (5, 6) is an independent favorable prognostic factor for GBM patients (3), but this marker does not fully account for the variability in survival (3), and additional markers are needed to be used for the prognosis of individual patients.
Homeobox genes encode transcription factors important for antero-posterior patterning during embryogenesis. In humans, there are 39 Class I homeobox (HOX) genes found in four genomic clusters (HOXA at 7p15.3, HOXB at 17q21.3, HOXC at 12q13.3, and HOXD at 2q31). Their spatial and temporal co-linear expression pattern is critical for body patterning during development (7, 8), and HOX genes also have critical post-developmental regulatory functions (9-12). Aberrant HOX gene expression initiates leukemias (13, 14), contributes to multiple types of solid tumors (15, 16), and has been detected in malignant astrocytomas (17). Furthermore, HOXA9 and MEIS1 may be part of a switch that regulates progenitor abundance by suppressing differentiation and maintaining self-renewal during myelopoiesis (18). Thus, among the 39 HOX genes, HOXA9 is of particular importance in leukemia and potentially other cancers.
In the present study, we identify a subgroup of GBMs with chromosomal domains of aberrant HOX gene activation, demonstrate a mechanism by which this domain of activation is reversible, define an association between HOXA9 expression and poor clinical outcome, and provide functional data on the effects of HOXA9 in cell proliferation and apoptosis.
Human tissue samples were obtained from the Neurosurgery Tissue Banks at UCSF and MDA, and would have been otherwise discarded. Tumor samples were examined by a neuropathologist to ensure that >90% of the tissue represented tumor, and classified according to current WHO guidelines (19). All samples were collected with informed consent, and investigations utilizing human tissues and clinical data were approved by the institutional review boards of UCSF and MDA. Initial exploratory survival analysis was performed on a set of 9 GBMs from UCSF. Validation of a trend between HOXA9 expression and poor survival was performed using two sets of expression array data (Affymetrix U133 arrays or Affymetrix U95Av2 arrays), one set consisting of 37 GBM samples from UCSF (20) and a second set composed of 63 GBMs from MDA (21) (Gene Expression Omnibus accession number GSE4271). Combined assessment of HOXA9 expression and MGMT promoter methylation was performed in 35 of the 37 GBMs from UCSF. Details of the tumor types and non-tumor samples are included in the online SI Materials and Methods. The nine GBM cell lines (A172, LN18, LN215, LN229, LN235, LN319, U87-MG, U178, and U373), and immortalized human astrocytes (hTERT/E6/E7) were maintained in DMEM with 10% FBS, the neurospheres were maintained in neurobasal medium supplemented with basic FGF and EGF (21), and the orthotopic xenografed GBMs were derived as previously described (22-24).
RNA isolation, cDNA synthesis, reverse-transcriptase PCR (RT-PCR), quantitative real-time PCR (qPCR), and chromatin immunoprecipitation (ChIP) analyses were performed by standard techniques in cell lines (SI Materials and Methods). DNA isolation, bisulfite treatment, and analyses of MGMT promoter methylation in primary tumor tissues from UCSF were performed as described previously (3, 25) and are detailed in SI Materials and Methods. A172 cells and neurospheres were treated with a PI3K inhibitor (LY294002) and rapamycin to investigate the mechanism regulating HOXA gene expression in GBM. Overexpression of HOXA9 by retroviral infection and siRNA-mediated silencing of HOXA9 were used to investigate the functional relevance of HOXA9 expression, which was assessed by cell proliferation and apoptosis analyses (SI Materials and Methods).
The Kaplan-Meier method was used to estimate overall survival (OS) and progression-free survival (PFS), where OS was measured from the time of surgical resection to death or the last date when the patient was known to be alive, and PFS was defined as the time from surgical resection to the time of demonstrated tumor growth on follow-up imaging or death, if death occurred before documented progression. Multivariate survival analyses by Cox proportional hazards models (backward selection) were performed to adjust for the effects of patient age and KPS. A P-value <0.05 was considered significant.
The 2-sided Student’s t-test was used to assess statistical differences in qPCR data, and in vitro cell death and apoptosis experiments. A repeated measures ANOVA was used to assess differences in the cell proliferation curves.
To identify samples that express a given gene and assign the confidence to each sample, we used model-based clustering to identify the mixture components. Analysis of gene expression microarray data, generation of heat maps for the HOX clusters, and investigation of correlations between expression of HOXA9 and other transcripts are detailed in SI Materials and Methods.
Initially, we assessed HOXA gene expression in normal brain tissue, primary diffuse glioma tissue, and GBM-derived cell lines using RT-PCR. Most HOXA genes were not expressed in fetal and adult non-tumoral brain tissues (Fig. 1A). Similarly, of eight primary grade 2 and grade 3 gliomas, HOXA gene expression was evident in only two cases (Fig. 1A). In contrast, all primary GBMs and cell lines demonstrated expression of multiple HOXA genes (Fig. 1A). Supplementary Fig. S1A shows representative RT-PCR analyses of HOXA gene expression in human astrocytoma samples, exemplifying tumors that were positive or negative for expression of specific HOXA genes. Collectively, these data suggest that aberrant HOXA gene expression is a common feature of the most malignant gliomas, GBM, and infrequent in lower-grade gliomas.
HOXA9 was the only HOXA gene whose expression showed a trend to associate with shorter survival of these 9 GBM patients (median survival of 9 versus 53 weeks, P=0.06). Supplementary Fig. S1B and S1C suggest HOXA9 expression in GBMs is not simply a reflection of the expression pattern of the tumor cell(s) of origin, nor a consequence of therapeutic agents, but rather represents an aberrant gene activation event associated with tumor progression.
We next analyzed expression of HOXA genes in two independent validation sets of expression array data, one set of 37 GBMs from UCSF (20) and a second set of 63 GBMs from MDA (21). To investigate whether aberrant HOXA activation was an isolated event, or part of a larger chromosomal mechanism, we created heat maps for the HOXA cluster and surrounding genes within 1Mb in each direction (Fig. 1B). In both sets, the HOXA cluster showed a distinct chromosomal domain of activation such that many of the HOXA genes were concurrently activated, while genes outside of the domain were mostly silent (Fig. 1B). Similar domains of transcriptional activation were also present in the HOXB, HOXC and HOXD clusters (Supplementary Fig. S2 and Fig. S3), suggesting a common, coordinated activation of these domains.
Heat map analyses reflecting the expression pattern across all four HOX clusters in UCSF and MDA tumor sets analyzed by expression arrays (Supplementary Fig. S4) revealed that (i) a subset of GBMs demonstrates widespread activation of HOX genes; (ii) the pattern of HOX activation is aberrant and does not resemble the coordinated co-linear expression observed during normal embryonic development; (iii) HOXA9 substantially contributes to the clustering into two sample groups in both datasets. To quantify which of the HOX genes account for the clustering, the fold change between the two groups for each gene was computed using the median value of the genes in each group (Table S1). The HOX genes with statistically significant different expression between the two heat map cluster groups in both the UCSF and MDA tumors are HOXA1-A5, HOXA7, HOXA9, HOXA10, HOXB7 and HOXC6.
We used the ONCOMINE database to investigate additional GBM expression array datasets for activation of the HOXA domain (26). A subset of GBM patients showed a similar profile of activation of HOXA genes in a study by Sun et al (27). Expression of most HOXA genes, with the exception of HOXA6 and HOXA13, showed a high positive correlation with HOXA9 expression (Table S2). HOXA11, which did not demonstrate a statistically significant positive correlation with HOXA9 expression in our tumor sets, also demonstrated a substantially lower correlation coefficient in the samples from Sun et al (27). Since HOXA6 and HOXA13 do not correlate with HOXA9 expression in all 3 tested tumor sets, a mechanism of chromosomal amplification encompassing the entire HOXA locus may not adequately explain the presence of such transcriptionally-active domains. Furthermore, we confirmed in the UCSF set that aberrant HOXA gene activation may be enhanced by, but is not reliant on, increased chromosome 7p15.3 copy number (P=0.5, data not shown).
In HeLa cells, the phosphoinositide 3-kinase (PI3K) signaling pathway can regulate EZH2 activity (28), a key component of the polycomb repressor complex 2 that promotes gene silencing by trimethylation of histone H3 lysine 27 (H3K27-3met) on target genes. AKT-mediated phosphorylation of EZH2 inhibits its histone methyltransferase activity, resulting in gene reactivation. Given that HOXA genes are known targets of EZH2-mediated methylation in some normal tissues (29), and that PTEN, a regulator of PI3K activity, is frequently inactivated in GBM (30, 31), we hypothesized that alterations in the PI3K pathway may be upstream effectors of aberrant HOXA gene expression in GBM. We first investigated a potential relationship between PI3K signaling and HOXA9 expression in A172 cells, a GBM cell line with PTEN homozygous deletion (32) and HOXA9 activation. Consistent with the PI3K pathway having a regulatory function on HOXA9 expression, HOXA9 transcript levels were significantly reduced in A172 cells following treatment with the PI3K inhibitor LY294002 (Supplementary Fig. S5A). This regulatory mechanism was reversible as HOXA9 was restored to pre-treatment levels 32 hours after A172 cells were placed in fresh medium lacking LY294002 (Supplementary Fig. S5A). The levels of HOXA9 protein were also decreased in LY294002-treated cells at 24h (Supplementary Fig. S5B). Considering the concomitant activation of several HOXA genes within domains of activation in primary GBMs, we hypothesized that HOXA genes in GBM are activated by a shared mechanism of transcriptional regulation. To address this question, the entire HOXA cluster was tested for suppression by LY294002 treatment of A172 cells. Expression of most HOXA genes was suppressed (range 39-74%) following treatment with LY294002, except for HOXA6 and HOXA13 (Fig. 2A); these two HOXA genes also did not correlate with HOXA9 expression in our primary tumor sets (Table S1) or in the cohort studied by Sun et al (27). These data suggest that HOXA9 expression is reversibly regulatable by the PI3K pathway in glioma cells and further suggest that this regulation extends throughout most of the 100+ kb chromosomal domain containing the HOXA cluster.
To further support our data implicating the PI3K pathway as a critical regulatory mechanism of HOXA gene expression, we next searched the Connectivity Map dataset for drugs that induce gene expression signatures involving HOXA9-associated genes (33). This analysis revealed common PI3K inhibitors, including LY294002 and Wortmannin, as top hits significantly associated with the HOXA9-derived gene expression signature (Table S3 and Supplementary Information). Additionally, in primary GBM xenografts derived from 18 patients, aberrant activation of HOXA9 was significantly associated with PTEN gene inactivation. PTEN mutations and homozygous deletion were previously examined by sequence analysis and multiplex PCR analysis, respectively (22, 24), further supporting the relevance of PI3K signaling for the regulation of HOXA expression (Table S4 and Supplementary Results - Aberrant expression of HOXA9 is associated with PTEN gene inactivation in primary GBM xenografts).
Therapeutic agents that inhibit mTOR, a downstream mediator of the PI3K pathway, are currently in clinical trials for GBM. To determine whether mTOR is a critical regulator of HOXA gene expression, A172 cells were treated with rapamycin, an inhibitor of mTOR activity, and the expression levels of the HOXA genes were assessed by RT-PCR. Rapamycin treatment resulted in a more modest inhibition of HOXA transcripts levels (Fig. 2B), suggesting that the mechanism by which the PI3K pathway regulates HOXA genes expression is primarily independent of mTOR. HOXA9 protein levels were also unaffected by rapamycin treatment as indicated by immunoblotting analysis (Supplementary Fig. S5C).
To determine whether the PI3K-mediated regulation of HOXA genes is also observed in other GBM cell lines, we tested how LY294002 treatment affected HOXA9 levels by qPCR in 2 sublines of a primary GBM grown in neurosphere conditions, both of which presented detectable levels of endogenous HOXA9. Similarly to the observations in A172 cells, the inhibition of PI3K in these neurospheres also repressed HOXA9 mRNA levels (Supplementary Fig. S5D), while the relationship between rapamycin and HOXA9 expression in the very small number of neurosphere lines is not as consistent (Supplementary Fig. S5E).
To elucidate the downstream effectors of PI3K-mediated regulation of HOXA gene expression, and considering that AKT mediates EZH2 histone methyltransferase activity in HeLa cells, we assessed whether the regulatory role of the PI3K pathway on HOXA9 expression could be linked to histone modifications. Chromatin immunoprecipitation (ChIP) analysis was performed on A172 cells that were either untreated or treated with LY294002 for 24 hours. A172 cells treated with LY294002 demonstrated reduced expression of HOXA9, as well as reduced H3K4-trimethylation (a mark of active chromatin) and concomitant increased H3K27-trimethylation (a mark of repressive chromatin) in 3 different regions near the HOXA9 promoter (Fig. 2C). When LY294002-treated cells were incubated for an additional 48 hours in fresh media without drug, HOXA9 expression, H3K27-trimethylation and H3K4-trimethylation returned to pretreatment levels (data not shown). Collectively, these data indicate that the regulatory role of the PI3K pathway on HOXA9 expression is linked to epigenetic histone modifications that appear to be fully and expediently reversible in GBM cells.
In order to test the functional consequences of HOXA9 overexpression in vitro, a GBM cell line (U87-MG) and immortalized human astrocytes (hTERT/E6/E7) stably overexpressing HOXA9 were established by retroviral infection and compared to their counterparts infected with control vector, which do not overexpress HOXA9 (Fig. 3A). The use of these two GBM models allowed us to test the effect of HOXA9 overexpression in tumoral (U87-MG) and immortalized but non-tumoral (hTERT/E6/E7) backgrounds. Expression of exogenous HOXA9 was associated with a modest but consistent increase in proliferation of hTERT/E6/E7 cells (Fig. 3B; P<0.001) and a decrease in their spontaneous apoptosis (Fig. 3C; P=0.032) compared to the parental cells. U87-MG cells stably expressing exogenous HOXA9 consistently showed a modest increase in proliferation, but only when the cells became more confluent (Fig. 3B; P<0.001). HOXA9 expression in U87-MG cells also decreased TRAIL-induced apoptosis (Fig. 3C; P=0.009). These data suggest that HOXA9 serves an anti-apoptotic role in immortalized human astrocytes and GBM cells, which may impact cell number/proliferation. Conversely, the inhibition of endogenous HOXA9 expression in A172 cells with siRNA (Fig. 3D) resulted in increased spontaneous apoptosis (Fig. 3D), further supporting an anti-apoptotic role of HOXA9 in GBM cells.
Based on the association of HOXA9 with poor prognosis in the initial test set of 9 GBM patients, we then performed survival analysis using the UCSF and MDA sets of expression array data (21). To validate the expression array data, RT-PCR for HOXA9 was performed on the subsets of cases with the highest and lowest HOXA9 expression array values, and the results were concordant (Fig. 4A-B insets). In the UCSF cohort, GBM patients whose tumors expressed HOXA9 (5 of 37, 14%) had a significantly shorter OS (median=22 weeks, CI=5-38 weeks) compared with those whose tumors lacked expression of HOXA9 (median=68 weeks, CI=36-100 weeks, P<0.0001, Fig. 4A). Similarly, in the MDA set, patients whose tumors expressed HOXA9 (14 of 63, 22%) had a shorter median OS (56 weeks, CI=10-102 weeks) than those whose tumors lacked HOXA9 expression (median OS=91 weeks, CI=46-136 weeks, P=0.03, Fig. 4B). Patients whose tumors expressed HOXA9 also had significantly shorter PFS. Median PFS was 4 weeks (CI=4-5 weeks) and 30 weeks (CI=22-38 weeks) in the UCSF set, and 46 weeks (CI=17-76 weeks) and 77 weeks (CI=22-132 weeks) in the MDA set for patients whose tumors did and did not express HOXA9, respectively (P<0.0001, Fig. 4C, UCSF set; P=0.03, Fig. 4D, MDA set). These associations remained statistically significant after adjusting for the effects of patient age and KPS in the UCSF set (P=0.006 for OS; P=0.01 for PFS). KPS data were not available for the MDA cohort.
MGMT promoter methylation in tumor tissue is currently the most powerful molecular prognostic indicator of favorable prognosis in GBM (3). However, a significant subset of patients whose tumors demonstrate MGMT promoter methylation do not experience survival benefit. To contrast the prognostic significance of HOXA9 expression with this biomarker standard, we performed methylation-specific PCR (MSP) analysis on DNA available from 34 of the 37 UCSF GBMs previously analyzed by expression array (Fig. 1B) and from the original 9 GBMs analyzed by RT-PCR (Fig. 1A). No DNA was available from the MDA set to perform MSP. MGMT promoter methylation was identified in 24 (56%) of the 43 cases (Supplementary Fig. S7); as expected (3), these patients demonstrated trends toward longer OS (P=0.1, Fig. 5A) and PFS (P=0.06, Supplementary Fig. S6A) compared with those whose tumors lacked MGMT methylation. HOXA9 and MGMT in combination showed statistically significant associations with OS and PFS. Patients whose tumors demonstrated HOXA9 expression and/or lack of MGMT promoter methylation had significantly shorter OS (median=31 weeks, CI=3-60 weeks) and PFS (median=13 weeks, CI =5-20 weeks) compared with patients whose tumors lacked HOXA9 expression and had MGMT promoter methylation (median=82 weeks, CI=41-123 weeks for OS, P=0.002, Fig. 5B; median=33 weeks, CI=22-44 weeks for PFS, P=0.001, Supplementary Fig. S6B). These associations remained statistically significant after adjusting for the effects of patient age and KPS (P=0.02 for OS; P=0.03 for PFS). Expression of HOXA9 was identified in five (21%) of the tumors with a methylated MGMT promoter and, among these patients, was associated with a significantly shorter OS (median survival of 9 versus 98 weeks, P<0.0001, Fig. 5C) and shorter PFS (median PFS of 4 versus 33 weeks, P<0.0001; Supplementary Fig. S6C).
Multiple HOX genes have been shown to be overexpressed in GBM cell lines and primary astrocytomas (17), suggesting a role for these genes in gliomagenesis. However, the mechanisms underlying HOX activation, in addition to their functional relevance in GBM cells, have not been explored. Our results define PI3K-regulatable chromosomal domains of transcriptional activation as a means of aberrant gene expression in cancer. This adds to a growing perspective on chromosomal domains of epigenetic silencing in normal and malignant cells (34-36). The domains of HOXA transcriptional activation we discovered have notable features: (i) they involve an important set of genes, which have critical roles in embryonic development, normal adult tissue function, and oncogenesis; (ii) they are mediated through reversible epigenetic histone modifications that are regulatable through the PI3K pathway, a pathway known to be frequently altered in many cancers (37); and (iii) they include aberrant expression of the oncogenic HOXA9 (13, 38), associated with histologic malignant progression, shorter time to tumor progression and shorter OS in GBM patients. Our functional data provide further support for an oncogenic role of HOXA9 in GBM. Furthermore, a recent study suggested HOX genes may be part of a glioma stem cell signature with prognostic significance in GBM patients treated with concomitant chemoradiotherapy (16). Our data identify PI3K pathway dysfunction as a potential driver of this signature, and suggests the exciting possibility that PI3K inhibitors may allow therapeutic reversal of the core element of this cancer stem cell signature.
We implicate the PI3K pathway in the generation of aberrant HOXA domains of activation in GBM. Activation of the PI3K pathway in glioma is significantly associated with increasing tumor grade, decreased apoptosis and adverse clinical outcome (37). In addition, PTEN, a central regulator of the PI3K pathway, is frequently altered in GBM and, when altered in lower-grade gliomas, portends a dismal prognosis (1). Although multiple downstream mediators of the PI3K pathway have been described, and much focus is placed on translational effects of PI3K pathway dysregulation, we propose that the pathway arm involving EZH2, a central member of the polycomb repressive complexes with intrinsic histone methyltransferase activity (39), mediates the PI3K-dependent transcriptional regulation of HOXA9, and potentially the whole HOXA cluster and many other genes. Akt-mediated phosphorylation of EZH2 suppresses trimethylation of lysine 27 in histone H3 (H3K27) by interfering with the ability of the EZH2 complex to interact with histone H3, leading to derepression of silenced genes (28). EZH2 can increase cancer cell proliferation when overexpressed (40), is associated with metastasis in prostate cancer (41), and is expressed at significantly higher levels in GBM tissue versus noncancerous brain tissue (26, 27) and low-grade gliomas (26, 42). Furthermore, EZH2 itself was among the genes that demonstrated co-expression with HOXA9 based on expression array data from both UCSF and MDA tumor sets. However, the activity of overexpressed EZH2 may well be inhibited by Akt-mediated phosphorylation. Regulation of HOXA expression is even more complex, involving other histone marks, non-coding RNA, and miRNA mediated regulation depending on the cellular context (43). Because epigenetic mechanisms are interdependent, PI3K-driven changes in H3K27 methylation may influence additional histone modifications, either directly or indirectly. In fact, the PI3K pathway has been linked to other specific histone modifications, such as H3K9 demethylation (44), acetylation of H3K9 and H3K18 (45), and deacetylation of H3K14 (45). Thus, therapeutic agents targeting components of the PI3K pathway may indirectly affect the histone marks of tumor cells. We hypothesize that particular profiles of histone modifications, in conjunction with the status of the PI3K pathway, may influence therapeutic decisions in the future. Indeed, while the histone profile of malignant cells is not currently used to direct therapeutic decisions, histone modifications are strongly correlated with clinical outcome of tumor patients (46-50).
Considerable focus has been placed on developing chemotherapeutic agents to inhibit mTOR, a downstream effector of the PI3K pathway that is an important regulator of cell growth and metabolism (51). While these agents are expected to suppress the growth effects of the mTOR pathway, it is less clear whether mTOR is a factor regulating the expression of HOXA genes in GBM. If these domains of transcriptional activation contribute to tumor growth in vivo, mTOR inhibitors alone may be suboptimal for improving the outcome in a subset of GBM patients. PI-103, a recently reported agent that dually inhibits PI3K and mTOR in glioma cells (52), or similar agents, may prove more efficacious.
MGMT promoter methylation is a favorable prognostic indicator in GBM, but is limited in its clinical applicability to individual patients. HOXA9 expression is a marker of poor prognosis, independent of MGMT promoter methylation status, and identifies a subset of the patients who, despite apparently favorable MGMT promoter methylation, have poor outcomes. Thus, the combined assessment of HOXA9 expression and MGMT methylation status in GBM may prove to be a promising effective prognostic tool. The grade specificity of HOXA9 expression according to histology and the association with poor outcome among GBM patients suggest that it may be useful as part of a molecular-based classification of gliomas. The prognostic relevance of HOXA9 expression in new experimental therapies warrants further investigation.
The authors would like to extend their appreciation to Drs. Corey Largmann and H. Jeffrey Lawrence for helpful discussion regarding HOXA9 and for sharing retroviral constructs, to Dr. Russell Pieper for supplying hTERT/E6/E7 cells, to Dr. Amith Panner for assistance with molecular techniques, to Dr. Susan Chang for helpful discussion regarding the clinical management of GBM patients, and to Drs. Scott Vandenberg and Andrew Bollen for neuropathological assessment of tumors used in this study.
Grant support: Funding agencies were National Institutes of Health (NIH CA094971 to J.F.C., NIH/NCI F32 CA113039-01 to J.S.S., Brain Tumor Spore (UCSF) grant CA097257 to C.D.J.); Karen Osney Brownstein (Endowed Chair to J.F.C.); UC Discovery Grant (Bio05-10501 to J.F.C and H.S.P.); Portuguese Science and Technology Foundation (SFRH/BD/15258/2004 to B.M.C.); Luso-American Development Foundation, Portugal (186/06 to B.M.C.).
Note: Supplementary data for this article are available at Cancer Research Online, including Supplementary Materials and Methods, and Supplementary Results.