Identification of novel vascular markers through gene expression profiling of tumor-derived endothelium

Background Targeting tumor angiogenesis and vasculature is a promising strategy for the inhibition of tumor growth and dissemination. Evidence suggests that tumor vasculature expresses unique markers that distinguish it from normal vasculature. Our efforts focused on the molecular characterization of endothelial cells (EC) in the search for selective markers of tumor vasculature that might be helpful for the development of effective therapeutic approaches. Results We investigated by microarray analysis the gene expression profiles of EC purified and cultured from tumor (ovarian carcinoma [HOC-EC]) and normal (human adrenal gland [HA-EC]) tissue specimens. We found distinct transcriptional features characterizing the EC of different origin, and identified 158 transcripts highly expressed by HOC-EC. We analyzed four of these genes, ADAM23, FAP, GPNMB and PRSS3, which were not previously known to be expressed by endothelium. In vitro experiments confirmed the higher expression of the selected genes in tumor-derived endothelium with no expression in tumor cells. In vivo investigation by in situ hybridization established that ADAM23, GPNMB and PRSS3 expression is localized on blood vessels of human cancer specimens. Conclusion These findings elucidate some of the molecular features of the tumor endothelium. Comparative transcriptomic analysis allowed us to determine molecular differences of tumor and normal tissue-derived endothelium and to identify novel markers that might be exploited to selectively target tumor vasculature.


Background
Interference with tumor vessel formation and blood supply has become a well-recognized approach of cancer therapy [1], as epitomized by the recent FDA approval of antiangiogenic drugs such as the humanized anti Vascular Endothelial Growth Factor (VEGF) antibody Avastin ® [2].
Vascularization is required for tumor growth and metastasis, and constitutes an important step in the control of cancer progression [3]. Experimental evidence correlates tumor vascularization with high malignancy and poor prognosis, and shows that elevated levels of angiogenic factors, such as VEGF and basic Fibroblast Growth Factor (FGF-2), are associated with tumor progression [4]. Moreover, expansive growth of metastasis appears to be linked to the ability to stimulate endothelial cell growth [5].
Tumor-induced vessels are ultrastructurally abnormal and differ from normal vasculature. They lack functional pericytes and are unusually permeable [6]. These abnormalities reflect the pathological nature of their induction, and underpin the novel therapeutic strategies directed against the vascular elements of the tumor stroma to selectively affect tumor vasculature and inhibit tumor growth [7].
In recent years an increasing body of evidence suggests that tumor vasculature expresses unique markers that distinguish it from normal vasculature. Numerous efforts have been aimed at the molecular characterization of tumor associated endothelial cells (EC) in the search for differences between tumor and normal tissue EC. A variety of techniques have been applied to study EC in vivo, isolated from ex vivo specimens or cultured in vitro. The in vivo phage display technology was used to identify molecules selectively expressed on tumor endothelium [8,9], while other approaches exploited the in vivo proteomic analysis to detect antigens accessible from the vasculature [10,11]. These techniques have allowed, mostly through studies conducted in experimental tumor models, the identification of vascular-specific motifs expressed by different tumor types and during different stages of carcinogenesis.
Gene expression of EC isolated from ex vivo human cancer and normal specimens, analyzed by means of Serial Analysis of Gene Expression (SAGE) or microarray, revealed differences at the transcriptional level in tumor and normal tissues [12][13][14][15]. In vitro models have been widely used to identify the specific functions of the endothelium and to understand the molecular modifications that might occur during angiogenesis. In these studies EC were exposed to different microenvironmental conditions, such as angiogenic growth factors and matrix proteins, shear stress, hypoxia, or tumor conditioned media [16][17][18][19][20]. All of these approaches led to the identification of mechanisms and molecules that are potentially involved in the formation of blood vessels. Despite these findings, attempts to discover tumor EC markers have always been hampered by technical difficulties in isolating functionally intact and phenotypically stable EC from tumor tissues. In fact, all the in vitro models have used cultures of immortalized EC and/or primary EC from human umbilical vein (HUVEC).
We developed a method to isolate and culture EC from tumor specimens (TdEC), by which we were able to show that TdEC in vitro maintain several of the features described for tumor vasculature and that they differ from EC isolated from normal tissue (i.e., from human adrenal gland HA-EC) [21]. These findings suggested that TdEC and HA-EC might represent useful tools to study tumor vessel properties and, ultimately, to identify tumor vasculature markers.
Here we describe the isolation of EC from human specimens and the characterization of the transcriptional profiles of EC derived from ovarian carcinoma (HOC-EC) and HA-EC by means of microarray technology. Through the comparison of HOC-EC and HA-EC gene expression we were able to demonstrate that tumor-associated EC differ from those in normal tissues, thus proving that molecular differences are maintained in vitro. We successfully identified molecules that are expressed selectively by the tumor vasculature in vivo, thus providing distinctive features to be exploited to selectively target tumor vasculature.

Characterization of Endothelial Cells (EC) isolated from human specimens
Endothelial cells were reproducibly isolated from human tissue specimens: HOC-EC from ovarian carcinoma, HKC-EC from kidney carcinoma, HA-EC from adrenal gland and HSk-EC from skin.
Before use, their endothelial origin was established. Endothelial cell cultures typically created a homogeneous monolayer of elongated cells in close contact and maintained an uniform morphology after few passages in culture; they formed capillary-like cord structures when plated on matrigel (data not shown and [21]). Their endothelial origin was established by positive immunostaining for platelet-endothelial cell adhesion molecule-1 (CD31/PECAM-1), von Willebrand Factor (vWF) and vascular E-cadherin (VE-cadherin) ( Table 1). In addition, no staining was observed when the cells were analyzed for the expression of alpha-Smooth Muscle Actin (alpha-SMA), a typical non-endothelial marker (Table 1). Flow cytometry studies assessed the ability to uptake low density lipoproteins (LDL), a characteristic of endothelial cells, and confirmed the positivity for CD31/PECAM-1 (data not shown).
The purity of the isolated EC was comparable to that of EC from umbilical vein (HUVEC), which were used as the reference population for EC (Table 1).

Genome wide transcriptional differences: tumor vs. normal tissue derived EC
Comparison of the transcriptional profiles of HOC-EC and HA-EC (experimental design and analyses detailed in Methods) showed that EC from tumor and normal tissue have distinct and characteristic expression patterns that are maintained in vitro. According to our selection criteria, 158 gene transcripts, corresponding to 179 probe sets, were more expressed by HOC-EC than by HA-EC. The genes belong to the different functional classes that are listed in Table 3.
Microarray results reveal an alteration in cell-cell and cellmatrix interaction as shown by the enhanced expression of molecules such as desmoglein 2, N-and OB-cadherin, protocadherin-9 and -17 and P-selectin (Table 3A). In keeping with the notion that the extracellular matrix in tumor differs from that in normal tissue [22] glypican-1, laminin alpha 3, tenascin C, versican, as well as different types of collagens (collagen typeI-alpha2, typeIII-alpha1, typeV-alpha1, typeXIII-alpha1 and typeXV-alpha1) and thrombospondin-2, were expressed at higher levels by HOC-EC than by HA-EC (Table 3A). Moreover, alterations of cytoskeletal protein involved in cell motility, for instance the actin and myosin related molecules MYOM1, MYL9 and RIPX, were also observed (Table 3A).
The peptidase family is the most represented among the enzymatic classes of genes expressed by HOC-EC (Table  3B). It includes carboxypeptidase A3, carboxypeptidase E, cathepsin Z, membrane metallo-endopeptidase, serinprotease 11, and urokinase-plasminogen activator. This is consistent with the concept that peptidases are very important for the activation of pro-enzymes or protein precursors, as well as in the degradation of the extracellular matrix [23].
Most of the listed genes are involved in signaling and comprise molecules previously shown to be involved in angiogenesis and other vascular functions, such as adrenomedullin receptor, CD44, endothelin-1 and -2, endothelin receptor-A and tissue factor (Table 3C).
Consistent with the notion that the vascular and nervous systems share similar signals and principles [24], 3 members of the GABA receptors family and several molecules involved in synaptic signaling, as well as elements involved in brain differentiation and axon elongation, were expressed at higher levels by HOC-EC (Table 3C).
Four members of the Solute Carrier Family (SLC1A1, SLC7A2, SLC16A4, SLC20A2) were over-expressed in HOC-EC, suggesting an alteration in the extra-intracellular exchange of metabolites (Table 3C). Among the transporters, the enhanced expression of the ATP-binding cassette family members ABCA2 and ABCB1 (Table 3B) should be emphasized. ABCB1, best known as MDR/pglycoprotein, is involved in drug removal from cells.
Among the class of cell cycle regulators, the high expression of the anti-apoptotic protein BCL2 is noteworthy (Table 3E).

Selected transcripts are expressed by tumor derived endothelial cells
In order to identify novel markers of tumor endothelium our attention focused on transcripts that are highly expressed by tumor-derived endothelium (i.e., HOC-EC), but whose products and functions were not previously reported as being involved in tumor angiogenesis or associated to vascular functions. Among them we distinguished four genes on the basis of their structural or functional characteristics, such as the presence of transmembrane domains and/or enzymatic activities: disintegrin and metalloproteinase domain 23 (ADAM23), fibroblast activation protein alpha (FAP), transmembrane glycoprotein nmb (GPNMB), and serine protease 3 or mesotrypsin (PRSS3). The relative expression in HOC-EC compared to HA-EC was 2.6 for ADAM23, 4.1 for FAP, 4.6 for GPNMB and 2.5 for PRSS3, as calculated from microarray data. These results were validated by real-time PCR in EC isolated from six neoplastic tissues (five ovarian and one kidney carcinoma specimens) and from four nonneoplastic tissues (three adrenal glands and one skin specimen) different from those investigated by microarray. These tumor or normal tissue derived EC are described in Table 1. Three HOC-EC and two HA-EC populations were investigated by Affymetrix GeneChip ® Human Genome U95A Arrays. This list of highly expressed endothelial genes was extracted from the microarray results by keywords search within the complete probeset list annotated using SOURCE batch search (see Methods for details).  List of the 158 gene transcipts more expressed by HOC-EC than HA-EC. Three HOC-EC and two HA-EC populations were investigated by Affymetrix GeneChip ® Human Genome U95A Arrays and analyzed searchingfor intrinsic differences, as described in Methods. To mimic the different environments in which EC might be embedded in vivo -either subjected to persistent stimulation imposed by the tumor cells or not, EC were investigated both in the presence ( Figure 1A) or the absence ( Figure 1B) of the angiogenic factors VEGF, FGF-2, of epidermal growth factor (EGF) and of fibronectin. Under both conditions the expression of ADAM23, FAP, GPNMB and PRSS3 genes was much higher in EC isolated from tumor specimens, as shown by the fold difference values ( Figures 1A and 1B). As specified by the p values, statistical analysis demonstrated a significant disparity in the expression levels for each of the genes being investigated ( Figures 1A and 1B).
These results confirm microarray findings and demonstrate that such a diversity is indeed an intrinsic characteristic associated to the tumor or normal tissue origin of the EC, regardless of the culture conditions. Sarcoglycan epsilon (SCGE), chosen as an internal control of a gene expressed to a much lesser extent by HOC-EC, confirmed the microarray results ( Figures 1A and 1B).

Expression in non endothelial cells
Having shown that ADAM23, FAP, GPNMB, and PRSS3 were expressed by a panel of EC, with the highest expression found in TdEC, we next investigated their expression in other cell types. The gene expression analysis revealed that they could be expressed to different extents also by other stromal cells, such as fibroblasts (HuFb and Malme3) and smooth muscle cells (UASMC), but in general not by tumor cells (Figure 2). Specifically, their expression in the carcinoma cell lines (1A9 and SKOV3 ovarian, MDA-MB-431 breast and HT29 colon) was not or was barely detectable (with the exception of PRSS3 by the HT29 colon carcinoma) (Figure 2). This lack of expression was supported by the Northern Blot results of a panel of tumor cell lines represented on the Multiple Tissue Expression array (Figure 3). Moreover, the Reference database for gene Expression Analysis (RefExA [25]) confirmed that the tumor cell lines did not show expression of the selected genes.
Conversely, the expression of FAP and GPNMB was higher in both the HuFb and MALME 3 fibroblasts than in EC, and FAP was also expressed at a higher level by the UASMC smooth muscle cells. ADAM23 and PRSS3 were instead expressed to a lesser extent by all the stromal cells analyzed (Figure 2).

Expression in healthy tissues
Tissue expression of ADAM23 and GPNMB was assessed using a Multiple Tissue Expression (MTE) array: a panel of human mRNAs in dot-blot format from 58 normal adult tissues, seven fetal tissues, and eight cell lines. As shown in Figure 3, ADAM23 and GPNMB were expressed by a limited subset of tissues, while tumor cells did not express either one of them. In particular, ADAM23 expression was highest in cardiovascular tissues and fetal heart, as well as in central nervous system (CNS) samples, while it was much lower in fetal brain (Figure 3). The expression of GPNMB was highest in cardiovascular tissues and fetal heart, as well as in placenta, while it was detected to a much lesser extent in lung, uterus, and bladder tissue. (Figure 3). These Northern Blot results for ADAM23 and GPNMB expression were confirmed by the RefExA database [25].   Data from RefExA [25] and literature report the expression of PRSS3 by brain, colon and pancreas [26], while FAP expression is described only in the activated stromal fibroblast of tumors [27].

Expression by tumor vasculature in vivo
In order to investigate the in vivo localization of ADAM23, GPNMB and PRSS3, human tumor tissues were analyzed by in situ hybridization.
Endothelial expression of the selected genes was confirmed in the tumor samples analyzed. As shown in Figure  4, the hybridization signals for ADAM23, GPNMB, and PRSS3 localized around the blood vessels, and exclusively in the vascular walls. According to the in vitro results, tumor cells did not express the transcripts, as shown by the absence of staining in the parenchyma of the tissues ( Figure 4A, B, C). Noteworthy was the finding that stromal compartments (except for the blood vessels) lacked the expression of either gene.
No staining was observed in normal brain and Alzheimer's-diseased brain ( Figures 4G, H, I). It is remarkable that ADAM23 and PRSS3 transcripts, whose expression in CNS was revealed by the MTE array and/or reported by RefExA [25], were not detected in normal or diseased brain, except for blood vessels in medulloblastoma or metastatic adenocarcinoma (Figure 4 and data not shown).
Sense probes did not hybridize with any of the tissue sections analyzed ( Figures 4D, E, F).

Discussion
Targeting the process of tumor neovascularization has become a promising anti-cancer strategy. This is particularly evident when vascular targeting agents are combined with conventional chemotherapy and radiotherapy, as demonstrated by recent reports and clinical trials [28]. However, the success of this approach depends on the availability of suitable endothelial markers that can be used to selectively target the tumor vascular compartment or to deliver a ligand-directed effector to the endothelium [7]. Herein, we successfully combined a method to isolate and culture endothelial cells (EC) from human specimens of tumor (TdEC) and non-neoplastic (NdEC) origin with oligonucleotide microarray analysis to compare their gene expression. We first showed that the TdEC transcriptional profile differs from that of NdEC, and, thereafter identified specific molecular features of tumor-derived endothelium.
Our laboratory previously established a procedure [29] to derive EC of high purity from human specimens, their origin assessed by immunohistochemical staining for the  three endothelial markers CD31/PECAM-1, VE-cadherin and vWF. Accordingly, microarray results confirmed their expression, but also revealed the expression of many other genes annotated as typically endothelial through SOURCE [30]. These included E-selectin, VEGFR-1, and VEGFR-2, further confirming our previous findings [21]. Despite being less successful with tumors (e.g., a 50% success rate out of 18 ovarian carcinoma specimens) than with non neoplastic (a 90% success rate for adrenal gland), the reproducible yield of EC from human specimens enabled us to explore molecular differences between tumor and normal endothelium.

Gene expression in different cell lines
The comparative transcriptomic analysis of EC from ovarian carcinoma (HOC-EC) and adrenal gland (HA-EC) revealed the higher expression of 158 gene transcripts by HOC-EC. The finding that tumor and normal tissue EC have distinct gene expression patterns concurs with other investigations, which describe genes that are preferentially expressed by endothelium from malignant cancer [12][13][14][15]. In these studies, SAGE [12][13][14] or microarray analyses [15] were executed on EC freshly obtained by immunoselection of ex vivo tissue samples. Endothelial cells in culture often lose their tissue-specificity and their specialized properties [31]. Our results, nonetheless, demonstrate that we were able to preserve in vitro some of the phenotypes that distinguish the diverse endothelia. Likewise, they confirm our previous findings [21] and agree with other reports [32,33] showing that, despite being in culture, EC isolated from different vascular beds possess characteristic gene expression profiles.
Here we describe for the first time a much higher expression of ADAM23, FAP, GNPNB and PRSS3 in EC derived In situ hybridization of human tissues Figure 4 In situ hybridization of human tissues. Shown are representative samples of medulloblastoma (A), brain metastasis of an adenocarcinoma (B) and Ewing sarcoma (C), hybridized with antisense probes for ADAM23, GPNMB and PRSS3, respectively. Positively stained blood vessels are indicated by arrows. Related hybridization with control sense probes for ADAM23, GPNMB and PRSS3 is shown in panel D-E-F. Normal brain samples, from different donors, hybridized respectively with antisense probes for ADAM23 (G) and GPNMB (H) are shown. Alzheimer's-diseased brain hybridized with antisense probe for PRSS3 is shown (I).
from six tumor samples than in EC from four normal tissue specimens examined by real-time PCR, whose origin differed from that investigated by microarray analysis. The differences we detected between TdEC and NdEC were maintained whether or not they were exposed to a combination of pro-angiogenic factors that presumably mimic the different environmental conditions under which EC might be embedded in vivo (that is, subjected or not to persistent stimulation imposed by the tumor cells, as in the case of blood vessels in adult normal tissues). These results suggest that the transcriptional differences are not culture condition artifacts, but, rather, true differences that distinguish EC of different origin (i.e., tumor vs. normal).
Despite having mainly utilized EC from ovarian cancer and from adrenal glands as the non-neoplastic counterpart, the results hold true for EC isolated from kidney cancer and skin specimens (even though only one EC population could be analyzed). These observations, would seem to indicate that the EC characteristics we observed are not organ-specific, but, rather, exclusive of TdEC. This is in agreement with a recent report [15] that compared EC freshly isolated from ovarian cancer and normal ovary in which several of the genes identified as distinctive of the tumor associated vasculature showed conserved expression across the EC from other tumor types, such as colon [12], brain [13] and breast [14].
Cultures of EC are valuable tools for biochemical and functional analysis, but they do not reflect the physiological situation entirely. All the information should be considered critically from the standpoint of what might occur in blood vessels in vivo. To this end, in situ hybridization results confirmed ADAM23, GNPNB and PRSS3 expression associated to blood vessels in tumor tissues, but the lack of their expression in cancer cells (as well as in normal tissue).
Overall, these findings support the premise that endothelial markers identified in vitro are likely to have in vivo relevance, and bear out the validity of our experimental setting designed to reproduce situations that might occur in vivo.
ADAM23 belongs to the cellular disintegrins, a family of membrane-anchored proteins that are potential regulators of cell-to-cell and cell-to-matrix interactions [34] and for which a specific binding interaction with alphaVbeta3 integrin has been demonstrated [35]. Our MTE array results showed that ADAM23 is expressed in fetal and adult heart, and also confirmed its previously reported expression by fetal and adult brain [34,35]. However, our in situ hybridization finding (lack of gene expression in normal or Alzheimer's-diseased brain) contradicted the latter, which implies that the expression of ADAM23 might indeed be much higher in tumor EC, including blood vessels in medulloblastoma, a highly malignant brain tumor, than in the brain. Further studies, aimed to produce antibodies to ADAM23, are necessary to prove its suitability as a tumor vascular target. Our results also revealed that ADAM23 was not expressed in cancer cells. This observation is consistent with the data available from RefExA and other recent reports [36], which describe hypermethylation of the ADAM23 promoter that silences its expression in breast and pancreatic cell lines and in the corresponding primary tumors. Our evidence of a high expression of ADAM23 by TdEC suggests that this protein might be important for establishing contacts and promoting adhesive functions during the pathological processes leading to the formation of new blood vessels within the tumor. In addition, it is intriguing to speculate that this protein -through its ability to bind alphaVbeta3 -could play some specialized role in the maintenance of vascular functions within the tumor. In keeping with this hypothesis is the fact that alphaVbeta3 is expressed by EC and has been shown to be involved in tumor neo-vascularization [37].
GPNMB is a type I membrane protein identical to the hematopoietic growth factor inducible neurokinin-1 type (HGFIN) [38] sharing similarities with neurokinin-1 (NK1), a receptor for tachykinins and a member of seven transmembrane G protein coupled receptors. HGFIN was shown to interact with Substance P (SP), which has been reported to induce EC proliferation and increase vascular density in an in vivo model of inflammation, and these effects could be inhibited by an NK1 receptor antagonist [39]. The action of SP via the NK1 receptor may be a direct effect, since NK1 receptors have previously been localized to endothelium [40], and both SP and selective NK1 agonists were shown to enhance in vitro EC proliferation [41]. Substance P in vivo also contributes to vascular permeability, plasma extravasation, and edema [42], all features that characterize the abnormal vasculature of tumors. Our evidence of a high expression of GPNMB by TdEC suggests that GPNMB may be a tumor stroma receptor for SP, involved in tumor angiogenic events and in determining the distinctive features of tumor vasculature. Corroborating this premise are data suggesting that, in addition to NK1 receptors, other mechanisms contribute in mediating in vivo SP effects on EC [39].
The PRSS3 gene encodes for a trypsinogen whose functions have not yet been fully elucidated [43]. PRSS3 transcript, whose expression in CNS is reported by RefExA, was not detected in normal or Alzheimer's-diseased brain, except for blood vessels in a brain metastasis. This finding might mean that the expression of PRSS3 is indeed characteristic of tumor EC, as seen for ADAM23. Our search in the RefExA and SOURCE databases revealed that PRSS3 was not expressed by tumor cells, with the exception of colon carcinoma. It seems very plausible that PRSS3 produced by TdEC may contribute to tumor angiogenesis, invasion and metastasis through its peptidase activity. This notion upholds evidence that the migration and invasion of EC in the tumor tissue requires the activation of pro-enzymes (such as pro-metalloproteinases) and protein precursors [44]. To this end, it is noteworthy that other peptidases (including urokinase-plasminogen activator [uPA] and FAP) were among the genes we found to be highly expressed by tumor-derived endothelium. In particular, uPA is capable of initiating the process of extracellular matrix degradation through the activation of plasminogen and matrix metalloproteases; in addition, it has been shown to be required for the migration of EC during the process of angiogenesis [45]. It is interesting that the expression of the uPA receptor (uPAR) was also seen to be higher in HOC-EC than in HA-EC (C. Ghilardi unpublished observation), a finding that would implicate the uPA-uPAR system in tumor driven angiogenesis.
Our results show that the cell surface protease FAP was also expressed at significantly higher levels by TdEC than by normal EC. FAP was originally reported to be an antigen recognized by F19 antibody in cultured fibroblasts [27]. Accordingly, its expression in our tests was found to be much higher in cultured fibroblasts (HuFb and MALME3) and in smooth muscle cells (UASMC). Nonetheless, while normal adult tissues are generally FAP-negative, its expression is detectable in the stroma of over 90% of carcinomas [27]. Such behavior suggested its suitability for stromal targeting and/or cancer detection and therapy, and in recent years has underpinned the development of a humanized anti-FAP antibody (sibrotuzumab) and the implementation of clinical trials [46]. Our results uphold the notion that FAP is a valid target for therapy, be it as stromal or as vascular marker. In keeping with our results, the in vivo expression of FAP on blood vessels was recently demonstrated by Ghersi et al. in invasive breast ductal carcinoma [47].
Many of the transcripts we found to be expressed at a high level by EC from neoplastic tissues were previously reported as being involved in tumor angiogenesis or expressed by tumor EC. In this regard, it is worth listing the various types of matrix proteins that we detected, namely: collagens alpha-2 type I, alpha-1 type III, V, XIII and XV, glypican-1, laminin alpha3, and tenascin C. Our results confirm similar findings from other laboratories [12,13,15] and denote a remodeling of the extracellular matrix known to promote blood vessel sprouting and growth. In keeping with the experimental data, and as an example, tenascin C was shown to be expressed in vivo and to colocalize with microvessels in the stroma of non-small cell lung cancer [48].
The expression of cell adhesion proteins was also altered in TdEC. This may be explained by the fact that migrating and proliferating EC must "loosen up", that is, detach and then create new interactions. Our results implicate CD44 and cadherins in this process, and confirm previous reports showing N-cadherin over-expression on proliferative state EC [49], and enhanced expression of CD44 on the vasculature of solid tumors compared to normal tissue [50]. Noteworthy is the finding that CD44 expression was up-regulated when the TdEC were exposed to angiogenic growth factors (C. Ghilardi unpublished observations), in accordance with previous results showing that FGF-2 and VEGF stimulate CD44 expression in EC [50].
Interestingly, ABCA2 and ABCB1 (MDR-1/p-glycoprotein 1) were also more abundantly expressed by TdEC, reflecting the previously reported higher expression of p-glycoprotein 1 in EC from glioma than from normal brain [51]. The ATP-binding cassette transporters (ABC) are responsible for the efflux of chemotherapeutic agents in cancer cells [52]. These data suggest that the multidrug resistance phenomenon in tumors may be due to both EC and tumor cells.
Molecules such as SMAD7, BAMBI and BCL2 are listed among the transcripts that are more highly expressed by HOC-EC than by HA-EC. SMAD7 and BAMBI interfere with the TGFβ pathway and may potentially block the anti-proliferative action of TGFβ [53]. Their expression may induce an enhanced proliferation of the TdEC, which is one of the steps of the angiogenic process. BCL2 is one of the most widely recognized anti-apoptotic factors, whose upregulation enhances EC survival and intratumoral angiogenesis, thus promoting tumor growth [54].
Emerging studies suggest that the vascular and the nervous systems share the same molecular signals in their development. The involvement of netrins, semaphorins, Robo/Slit and VEGF families in both angiogenesis and neurogenesis has been demonstrated in recent years [24]. In keeping with these findings, our results revealed that TdEC express numerous molecules that are associated with the CNS or are implicated in brain differentiation and axon guidance. In particular, ADORA2B was identified as a novel netrin-1 receptor and reported to mediate axon outgrowth [55]. In a similar way -and considering that netrin-1 is pro-angiogenic [56] -our results allow speculatation that ADORA2B could mediate the sprouting of vessels during tumor neovascularization.

Conclusion
Targeting angiogenesis and vasculature represents a promising option to control tumor growth and dissemination. To exploit this opportunity, however, selective targets in tumor associated vasculature need to be identified. We describe a model consisting of in vitro cultured endothelial cells isolated from human normal and neoplastic specimens that maintained their respective molecular phenotype and enabled us to determine molecular differences between tumor and normal tissue derived EC. Specifically, we identified four transcripts whose expression is much higher in endothelial cells from tumors than from normal tissue. The in vivo expression in cancer-associated blood vessels was confirmed by in situ hybridization of human specimens.
This work evinces some of the distinctive features of tumor-derived endothelium, and identifies potential markers that may provide the groundwork for novel therapeutic strategies. Future studies are needed to clarify the functions and roles of these molecules in tumor progression and vascularization.

Isolation, culture and characterization of endothelial cells (EC)
The method for isolation of EC has been described in detail by [29,21]. Briefly, tissue specimens were digested by type I collagenase (EC 3.4.24.3 -clostridiopeptidase A, Sigma-Aldrich, St. Louis MO, USA -Milan, Italy) and the suspension plated. Six to 10 days later, the EC were positively purified using anti-CD31 antibody-coated magnetic beads, seeded and cultured as detailed in the supplementary information [see Additional file 1].
Human umbilical vein EC (HUVEC) were obtained following the procedure described by Jaffe [57] and cultured as detailed in [58].

RNA isolation
Total RNA was isolated from cultured cells using Trizol ® (GIBCO-Invitrogen) or RNeasy Mini Kit (Qiagen, Maryland, USA) following the manufacturer's protocols. Potential genomic DNA contaminations were removed by DNAse (Ambion, Austin, USA) treatment followed by RNA CleanUp with RNeasy Mini Kit. The purity and integrity of the RNA was checked by gel electrophoresis, and concentration was determined spettrophotometrically.

Genome wide analysis of gene expression by microarray Transcriptional difference: HOC-EC vs. HA-EC
We are aware that culture conditions alter gene expression. To this end, all the EC under investigation were analyzed under two different conditions: either seeded onto type I collagen coated tissue culture plastic (BioCoat-Becton Dickinson, Bedford MA, USA) in endothelial cell basal medium (EBM -Clonetics-BioWhittaker, Walkersville MD, USA) plus 5% FBS or onto BioCoat plasticware additionally coated with 1 μg/cm 2 fibronectin (Plasma Fibronectin, BD-Biosciences, Bedford MA, USA) in EBM plus 5% FBS supplemented with and 10 ng/ml Epidermal Growth Factor (EGF), 10 ng/ml Vascular Endothelial Growth Factor (VEGF), 2 ng/ml Fibroblast Growth Factor (FGF-2) (all from R&D System, Minneapolis MN, USA); from here on respectively called "unexposed" or "exposed" to a tumor/angiogenic environment.
To maximize the likelihood of detecting the intrinsic differences between HOC-EC and HA-EC by minimizing the influence of the artificial environment upon them, we opted to select only those genes that were differentially expressed regardless of the culture conditions.

Microarray analysis
GeneChip ® Human Genome U95A Arrays (Affymetrix UK Ltd, High Wycombe, UK) that monitor the expression of approximately 12,000 genes were utilized to analyze the transcriptional profiles of 3 independent HOC-EC together with 2 HA-EC populations, under both "exposed" and "unexposed" culture conditions. Raw Affymetrix data were then processed for each of the 10 GeneChip ® arrays.

Data analysis
For each probeset, the mean of the average difference values for the two "exposed" HA-EC samples was calculated only if both were greater than 20 arbitrary units. Similarly, the mean of at least two average difference values for the three "exposed" HOC-EC samples was calculated only if each of them was greater than 20 arbitrary units. The same was done for the "unexposed" samples. Then, the log2 of the HOC-EC versus HA-EC ratios was calculated only for the probesets having both means available. Probesets with log-ratios greater than 1.3 (corresponding to a fold change of 2.5) for both "exposed" and "unexposed" samples were considered. Such a threshold was obtained by fitting the distributions of the "exposed" and "unexposed" log2-ratios with a normal distribution and calculating upper bound of the 99% confidence interval.
Gene annotation the probesets expressed to a greater extent in HOC-EC than in HA-EC were divided into functional classes using the publicly available Est Annotation and the Keyword Clustering Machines [60] and were then manually checked.

Typical endothelial genes: HOC-EC and HA-EC
Endothelial markers represented on the Affymetrix Gene-Chip ® were found by keyword search within the complete probeset list annotated using the publicly available SOURCE batch search [30,61]. The mean of average difference values for each transcript in HOC-EC and HA-EC was calculated. Genes with the mean value greater than 500 were included in the list of expressed endothelial markers shown in Table 2.
Additional information on gene expression was obtained from the publicly available Reference database for gene Expression Analysis (RefExA) of the Laboratory for System Biology and Medicine (LSBM) (RCAST, University of Tokyo) [25]. All assays were performed using duplicate samples of each cDNA synthesis. Reverse transcription reactions and realtime PCR analyses were carried out at least twice for each gene transcript to determine consistency of results.

Quantitative real-time PCR
Gene expression of all transcripts was normalized to the endogenous control gene human 18s rRNA and for each population the ΔCt value was calculated (ΔCt = Ct targetgene -Ct 18s ). Statistical analysis with Student's t-test (unpaired) on ΔCt values was performed to compare the expression levels of target genes in EC samples from normal and neoplastic tissues.
Fold differences were calculated according to the comparative ΔΔCt method (GeneAmp 5700 SDS User's Bulletin; Applied Biosystems).

Northern Blot
For the analysis of tissue distribution, the MTE Array 2 (Clontech, Palo Alto CA, USA) was hybridized following manufacturer's instructions. Briefly, the filter was prehybridized for 30 min at 65°C with ExpressHyb (Clontech). Hybridization was carried out O.N. at 65°C in ExpressHyb containing 130 μg/ml fish sperm DNA, 6 μg/ ml Cot -1 DNA and the 32 P-labeled cDNA probe. Filter was washed with SSC2x/SDS0.1% at 65°C, followed by washes at 55°C with SSC0.1x/SDS 0.5%. Because the panel of human mRNAs loaded onto the MTE Arrays in dot-blot format are normalized using eight different housekeeping genes, the hybridization results can be attributed to actual differences in target mRNA abundance.
Densitometric analysis of the dots was performed with Gel-Pro Analyzer Software (Media Cybernetics, Silver Spring, MD). Optical Density (OD) values are reported in Figure 3.
cDNA probe for ADAM23 corresponding to nt 2241-2743 of GeneBank:NM_003812, and cDNA probe for GPNMB corresponding to nt 416-984 of GeneBank:NM_002510 were PCR amplified and gel purified.

In situ hybridization
In situ hybridization experiments were carried out with a mixture of specific biotin-labeled oligonucleotide anti-sense (or sense) probes for ADAM23, GPNMB and PRSS3 (Metabion International AG, Martinsried, Germany) listed in the supplementary Table 2 [see Additional file 1]. Human cancer specimens, such as adenocarcinoma brain metastasis, angioma from a Von Hippel Lindau disease patient, Ewing sarcoma, and medulloblastoma, were analyzed along with normal brain and Alzheimer's-diseased brain.
After hybridization, sections were washed for 15 min in 2× SSC and for 10 min in water at room temperature. They were rinsed three times in PBS/0.1% Triton, and incubated for 20 min with blocking buffer (PBS/0.1% Triton, 5% FCS). Bound probes were detected using a mouse antibiotin antibody (DakoCytomation, Glostrup, Denmark). Amplification of the signal was achieved with the Vectastain elite ABC Kit (Vector Laboratories, Burlingame, CA) according to the manufacturer's protocol. Colorimetric detection was completed with the AEC Substrate kit (Vector Laboratories). Finally, the tissue sections were counterstained with hematoxylin and embedded on an aqueous mounting medium.

Ethical consent
Endothelial cells have been isolated from biopsies collected at the San Gerardo Hospital (Monza, Italy) from patients undergoing therapeutic surgery from September 1998 to March 2004. The collection and use of the tissue samples was approved by the Local Scientific Ethical Committee in compliance with the principles enunciated in the Helsinki Declaration [62]. Tissue samples (including tumors) used for in situ hybridization were acquired from the Oxford based Thomas Willis brain bank. The collection of the tissues was done in accordance with the requirements of the Human Tissue Act, with full informed consent of participants and relatives. Ethical approval for research to be conducted on the tissues was granted by COREC (No: 1656).