Identification of Genes Expressed in Malignant Cells That Promote Invasion

Cancer has been described as a progression of genetic mutations in an aberrant tissue mass, and many oncogenes and tumor suppressor genes have been identified (1). The notion that tumor progression after the genetic events that initiate malignancy critically involve interactions between the malignant cells and the normal cells in the tumor environment has come to the fore more recently (2, 3). It is now accepted that stromal cells as well as cells from distant locations actively contribute to tumor growth, invasion, and metastasis (4, 5, 6). Epithelial cells use a wide variety of factors to sense and to respond to their environment. These include cell adhesion molecules, such as intergrins (7), cadherins and the associated catenins (8), and growth factors, such as fibroblast growth factor (9), and transforming growth factors α and β (10).

Stroma often comprise a major portion of solid tumors such as breast, colon, and prostate carcinoma (11, 12, 13). The most prominent stromal cell type, the myofibroblast, secretes large amounts of extracellular matrix proteins responsible for the host desmoplastic response (14). Myofibroblasts express many growth factors involved in tumor angiogenesis including vascular endothelial growth factor, basic fibroblast growth factor, and transforming growth factor β, thus stimulating endothelial cell migration, invasion, proliferation, and vessel stability (15, 16). Several in vitro models have been developed to study tumor-stromal interactions. Janvier et al. (17) suspended a mixture of fibroblasts, endothelial cells, and human PC3 prostate carcinoma cells in collagen or fibrin and demonstrated the importance of cellular interactions between endothelial cells and fibroblasts in tube formation. Shekhar et al. (18) showed that breast carcinoma-derived fibroblasts, breast carcinoma epithelial cells, and human umbilical vascular endothelial cell cultured on a layer of Matrigel developed a compartmentalized spheroid with a core of fibroblasts and endothelial cells surrounded by epithelial cells. These heterogeneous spheroids had increased proliferation and invasion, degradation of extracellular matrix, and expression of matrix metalloproteinase 9. These cell culture models support the observation that stromal cells play an important role in tumor angiogenesis.

We have described previously a multicellular stromal invasion assay using three cell types involved in tumor angiogenesis (endothelial cells, myofibroblasts, and tumor cells; Ref. 19). SKOV3 ovarian tumor cells suspended in a collagen I matrix surrounded by Matrigel were cultured with fluorescently labeled mature endothelial cells or myofibroblasts or with a 1:1 mixed population of mature endothelial cells and myofibroblasts. The mature endothelial cells adhered to the outside of the tumor cell cluster, whereas myofibroblasts invaded the cluster and localized in the tumor cell cluster. The mixed population of mature endothelial cells and myofibroblasts colocalized in the center of the SKOV-3 tumor cluster. Thus, it appeared that the myofibroblasts enabled invasion by the mature endothelial cells.

Serial analysis of gene expression (SAGE) has been extensively used for expression analyses of various types of cancers for identifying novel diagnostic and prognostic markers and potential therapeutic targets, as well as for identifying pathways up-regulated during malignant transformation and progression (20, 21, 22, 23). Herein SAGE was used in combination with the tumor cell cluster stromal invasion assay with eight human tumor cell lines to systematically investigate the genes expressed by tumor cells that promoted interactions between tumor cells and stromal cells. Combining SAGE with the in vitro tumor cluster model to identify genes expressed by tumor cells that enabled stromal cell invasion provided a statistically robust gene expression analysis based on the eight human tumor cell lines. Several genes identified through the analysis were validated using quantitative, real-time PCR and in situ hybridization.

Human adult dermal microvascular endothelial cells (HMVECs) and EGM2-MV medium were purchased from Clonetics (Walkersville, MD). MDA-MB-231, [American Type Culture Collection (ATTC) no. HTB-26, human breast carcinoma cell line], A-375 (ATCC no. CRL-1619, human melanoma cell line), LNCaP (ATCC no. CRL-10995, human prostate carcinoma), A-549 (ATCC no. CCL-185, human lung carcinoma), PC-3 (ATCC no. CCL-1435, human prostate carcinoma cell line), DU-145 (ATCC no. HTB-81, human prostate carcinoma cell line), and SKOV-3 cells (ATCC no. HTB-77, human ovarian carcinoma cell line) were purchased from ATTC (Manassas, VA). The melanoma cell line, Mel624 was a gift from Dr. Steven Rosenberg (National Cancer Institute, Bethesda, MD). Human adult dermal fibroblasts (HDFs) were a gift from Dr. James Gailit (State University of New York, Stony Brook, NY). DMEM, fetal bovine serum, and F12/Ham’s medium were purchased from Life Technologies, Inc. (Gaithersburg, MD). Falcon tissue culture flasks, 24-well plates, and Matrigel were purchased from Becton Dickinson (Franklin Lakes, NJ). The fluorescent label, PKH26, was purchased from Sigma Chemical Company (St. Louis, MO). Collagen I (Vitrogen) was supplied by Cohesion Technologies (Palo Alto, CA). PCR primers were purchased from Integrated DNA Technologies (Coralville, IA). Trizol was purchased from Sigma, and the RNA Extraction kit was obtained from Qiagen (Valencia, CA). The High Capacity cDNA Archive kit, Taqman rRNA Control Reagents, Taqman Universal PCR Master Mix, and Sybr Green PCR Master Mix were purchased from Applied Biosystems (Foster City, CA). Tissue sections from a human ovarian metastatic tumor (Cooperative Human Tissue Network #25412) and a nonmetastatic tumor (Cooperative Human Tissue Network #25439) were obtained from the Cooperative Human Tissue Network at the National Cancer Institute (Bethesda, MD).

All of the cells were grown in a humidified incubator at 37°C and 5% CO₂. HMVECs (up to passage 9) were maintained in EGM2-MV. The eight human tumor cell lines and human dermal fibroblasts were grown in DMEM-supplemented 10% fetal bovine serum and 10 units/ml penicillin/10 μg/ml streptomycin (Invitrogen Life Technologies, Inc., Grand Island, NY). Human dermal fibroblast cultures at high passage number were confirmed to be substantially comprised of myofibroblasts as indicated by α-smooth muscle actin expression analyzed by fluorescence activated flow cytometry (data not shown; Ref. 11).

The in vitro tumor cell cluster model of stromal invasion has been described (19). Briefly, a thick layer of Matrigel (300 μl) was added to each well of a 24-well plate and allowed to polymerize. A plug of Matrigel of ∼1 mm in diameter was removed using a glass pipette under light vacuum. The resulting space was filled with tumor cells (1 × 10⁶) suspended in a 2.4 mg/ml collagen I solution (5 μl) prepared according to the manufacturer’s suggestions. The collagen was allowed to polymerize for 30 min. Myofibroblasts (HDF) were fluorescently labeled with PKH26 (red) according to the manufacturer’s suggested protocol. Briefly, myofibroblasts suspended in serum-free medium were incubated in the presence of 2.5 μm PKH67 diluted for 5 min. The labeling was terminated with 1 ml of fetal bovine serum for 1 min followed by three washes in serum-containing medium. After the washes, the labeled myofibroblasts were suspended in EGM2-MV and counted. HMVECs were infected with enhanced green fluorescent protein-expressing adenovirus at 200 multiplicity of infection for 18 h. A total of 30,000 cells (15,000 HMVECs and 15,000 myofibroblasts) were added to each well in EGM2-MV (1 ml). After 48 h, the fluorescently labeled cells were visualized using a fluorescein (enhanced green fluorescent protein/HMVEC) or rhodamine (PKH26/myofibroblasts) filter. In some experiments, 4′,6-diamidino-2-phenylindole-labeled human MDA-MB231 breast carcinoma cells and human PC-3 prostate carcinoma cells were used. Fluorescent and bright field images were captured with a ×4 objective on a Sony DXC-390 digital camera using Scion Image version 1.62c. Experiments were repeated three times.

SAGE libraries for the eight tumor cell lines were constructed as described (24). The SAGE libraries corresponding to each tumor cell line were retrieved from either the Genzyme proprietary database or the public Gene Expression Omnibus database.¹ Table 1 lists the sample and tag information for the libraries. Tag counts were normalized to 50,000 total library counts for each library. Initially 2 tag counts were removed from at least 2 of the 11 libraries as a filter for erroneous tags. Libraries were separated into the “Stromal Invasion” group and the “Non-Stromal Invasion” group based on the behavior of the cell lines in the in vitro assay. Two statistical tests were applied to the SAGE tag counts, t test for the group comparison, and χ² test for the comparison of the averages of each group. Confidence interval levels (90% for the χ² test and 95% for the t test) were used as significance filters. Hierarchical clustering was performed on filtered libraries using GeneSpring software release 5.0.2 build number 954 (Silicon Genetics, Redwood City, CA). Pearson correlation for similarity measurement and the minimum distance was set to 0.001. SAGE tags were mapped to UniGene (Build #157) clusters using modified tag to gene mapping. This mapping was generated to maximize both the specificity and coverage of the UniGene clusters that mapped to a single tag. UniGene cluster numbers were additionally used as links to obtain functional and cellular localization annotation for individual genes.

Each of the eight human tumor cell lines were grown to confluence in T75 flasks suspended and lysed using TRIzol. Cellular RNA was isolated by phenol:chloroform extraction followed by column isolation using the Qiagen RNA Extraction kit. cDNA was generated using the High Capacity cDNA Archive kit. Real-time PCR for was performed with Sybr Green PCR Master Mix using PCR primers on an ABI Prism 7900 Sequence Detection System (Applied Biosystems). Relative mRNA expression was determined by dividing the threshold of each sample by the threshold of 18S, solving for 2^x and adjusting the relative value to the whole number.

In situ hybridization was carried out using a modification of reported procedures (21). cDNA fragments for the BST2 and TNFRS5 mRNAs were generated by PCR amplification of fragments ranging from 200 bases to 650 bases, using primers with T7 promoters incorporated into the antisense primers. Digoxigenin riboprobes were generated by in vitro transcription in the presence of digoxigenin, according to manufacturer’s instructions (Roche, Indianapolis, IN). Human ovarian tumor sections were deparaffinized in xylene, washed in 100% ethanol, and then hydrated in 85%, 75%, and 50% ethanol in distilled water. After incubation in diethylpyrocarbonate-treated water, sections were permeabilized by treatment with pepsin in 0.2 n hydrochloric acid, washed briefly in PBS, then fixed in 4% paraformaldehyde. The sections were acetylated in acetic anhydride/0.1 m triethanolamine (pH. 8.0), equilibrated for 10 min in 5× SSC, and prehybridized for 1–2 h at 55°C in mRNA hybridization buffer (DAKO, Carpinteria, CA) and then hybridized with digoxigenin riboprobes (100–200 ng/ml) in mRNA hybridization buffer (DAKO) overnight at 55°C. After removing unbound riboprobes by washing, sections were incubated with RNase (Ambion, Austin, TX) to remove any nonspecifically bound riboprobe and treated with peroxidase block (DAKO) to eliminate any endogenous peroxidase, then blocked with a 1% blocking reagent (DIG nucleic acid detection kit; Roche), containing rabbit immunoglobulin fraction (DAKO) in Tris-buffered saline). Rabbit antidigoxigenin-horseradish peroxidase (DAKO) was used to detect the riboprobes and to catalyze the deposition of biotinylated tyramide (Gen-Point; DAKO) according to the manufacturer’s instructions. Final detection was accomplished through rabbit antibiotin-conjugated to alkaline phosphatase (DAKO). Alkaline phosphatase was visualized with Fast Red (DAKO) for 10–60 min at room temperature and then counterstained in hematoxylin. The nuclei are blued with ammonium hydroxide for 30 s, and then mounted with crystal-mount (BioMeda, Foster City, CA).

To determine whether tumor cell lines could be grouped by their ability to enable stromal invasion, eight human tumor cell lines were analyzed in the tumor cell cluster model for stromal invasion (19). Tumor cell clusters (1 × 10⁶ cells) suspended in collagen and surrounded by Matrigel were cultured in the presence of 1:1 mixtures of fluorescently labeled endothelial cells (HMVECs) and myofibroblasts (HDFs; Fig. 1). After 48 h, the wells were scored for fluorescence localization within the tumor cell cluster. Among the eight human tumor cell lines, the breast carcinoma cell line MDA-MB-231, the ovarian carcinoma cell line SKOV-3, and the melanoma cell lines A375 and Mel624 had the greatest concentration of fluorescence within the tumor cell cluster. However, in the wells containing tumor cell clusters of the non-small cell lung carcinoma cell line A549, and the prostate carcinoma cell lines LNCaP, PC-3, and DU145 most of the fluorescence remained outside of the tumor cell cluster. Tumor cell clusters that were efficiently invaded by the myofibroblasts and endothelial cells decreased in size in part possibly from the contraction of the collagen induced by the myofibroblasts and in part from migration of the tumor cells from the cluster (Figs. 1 and 2; Ref. 25). In a second experiment, human MD-MBA-231 breast carcinoma cells and human PC-3 prostate carcinoma cells were labeled with a tracer concentration of 4′,6-diamidino-2-phenylindole and were subjected to the same assay described above. After 48 h, fluorescent images showed the MDA-MB231 cells moving into the Matrigel, whereas the PC-3 cells remained in the collagen plug (Fig. 2).

SAGE data were used to obtain a comprehensive, unbiased comparison of gene expression between the human tumor cell lines that underwent efficient invasion by the myofibroblasts and endothelial cells with those cell lines that did not. In each SAGE library, sequence tags of 10–11 bases from the last CATG site at the 3′ end of each transcript and their respective counts were acquired. The relative abundance of a specific tag in a SAGE library was considered to be proportional to the expression level of its corresponding mRNA (24). There were total 64,478 unique SAGE tags generated from 10 SAGE libraries obtained from monolayer cultures of the eight human tumor cell lines. After normalization of each of the libraries to 50,000 SAGE tag counts, 7,791 tags had counts of ≥2 in at least 2 of the 10 libraries. To compare the expression profiles of the eight human tumor cell lines, an initial hierarchical clustering of all of the SAGE libraries was carried out using GeneSpring software (Fig. 3). There was no clear grouping in the clustering tree among those tumor cell lines that were efficiently invaded by the myofibroblasts and endothelial cells and those that were not. Thus, the expressed genes that allowed efficient stromal invasion were likely represented by a small percentage of the total transcripts in the tumor cells and differences in the expression levels of these genes from cell line to cell line were not sufficient to direct the clustering process. Tumor cell lines from the same tissue of origin were not always nearest neighbors and did not form distinct subclusters in the GeneSpring analysis. This type of observation has been made previously when analyzing endothelial cells from human tumors along with human umbilical vascular endothelial cell and HMVEC from cell culture, and when comparing normal cells and malignant cells from the same tissues of origin. Therefore, the mixing of different tumor cell lines in the clustering diagram was not surprising (20, 21).

The availability of 10 SAGE libraries representing the eight human tumor cell lines provides multiple samples in each phenotypic group allowing statistical methods to be applied to the data (26, 27). A combination of two statistical tests was used to extract tags with significant differential expression. t test, which has been widely used when comparing gene expression data from microarray experiments, was applied for group comparisons. The t test was not able to eliminate tags with very low counts in one group and zeros in the other group. Therefore the χ² test, a frequency-based test, was used as an additional filter. When significance values of P ≤ 0.05 for t test and 0.1 for c² test were applied, 99 SAGE tags emerged with significant expression differentials between the two groups of tumor cell lines. These SAGE tags corresponded to 75 known genes, 25 previously un-named genes, and 3 unmatched tags. When the 10 libraries were subjected to GeneSpring analysis using the 99 differentially expressed SAGE tags, two distinct groups corresponding to the efficiency of invasion by the stromal cells into the tumor clusters formed (Fig. 4 A).

Two of the SAGE libraries included in the statistical analysis had <20,000 tags and, thus, may have been artificially elevated or reduced when the tag counts were normalized to 50,000 (27). To evaluate the effect of library size on the analysis, a statistical comparison was carried out on the eight SAGE libraries that exceeded 20,000 tags. After applying the same statistical criteria described above, 87 SAGE tags emerged as significantly differentially expressed between tumor cell lines that were efficiently invaded by stromal cells and those that were not. These SAGE tags corresponded to 64 known genes, 21 previously un-named genes, and 5 unmatched tags. The eight libraries with >20,000 SAGE tags formed two clusters when subjected to GeneSpring analysis (Fig. 4 B). Forty-seven of the 87 SAGE tags from the eight library analysis overlapped with the 99 tags from the 10 library analysis.

To validate the differential gene expression profiles identified by analysis of the SAGE libraries, 30 genes were selected for real-time PCR examination in the eight tumor cell lines. The differential gene expression of 9 of the genes analyzed by real-time PCR showed good correlation with normalized SAGE tag counts. The strongest correlations were for bone marrow stromal antigen 2 (BST2or HM1.24), stathmin-like 3 (STMN3), tumor necrosis factor receptor superfamily member 5 (TNFRSF5), and hepatocyte growth factor tyrosine kinase substrate (HGSor HRS; Fig. 5). The frequency of the expression of BST2, STMN3, TNFRS5, and HGSin 150 SAGE libraries including 87 libraries representing malignant cell lines or tissues and 63 libraries representing normal cells or tissues was determined (Table 2). The most widely expressed gene was HGS, which was observed in ≥50% of 147 of the SAGE libraries, the exceptions being fibroblast cells, normal ovary, and renal cell carcinoma. The most highly tumor-selective gene expression was observed for BST2. BST2 was expressed by 100% of the brain tumors, renal cell carcinomas, hepatocellular carcinomas, and lymphomas, 75% of the pancreatic carcinomas, and >50% of the breast cancers, melanomas, and ovarian cancers. BST2 was also found frequently in normal kidney, liver, and in dendritic cells. The tumors that most frequently expressed all four of these genes were brain tumors, hepatocellular carcinoma, lymphomas, and pancreatic cancers. Normal kidney SAGE libraries also highly expressed these genes (STMN3, 100%; TNFRS5, 85%; BST2, 50%; and HGS, 50%). The other normal tissues that had relatively high expression of these representative genes were liver and dendritic cells. The 29 endothelial cell SAGE libraries includes both tumor endothelial cells and normal endothelial cells, and all four of the genes had some level of expression in the endothelial cell libraries ranging from 90% for HGSto 10% for TNFRS5.

In situ hybridization for BST2 and STMN3 was performed on samples of metastatic and nonmetastatic ovarian cancer (Fig. 6). Substantially higher levels of both BST2 and STMN3 were observed in the metastatic ovarian cancer tissue than the nonmetastatic ovarian cancer tissue.

The tumor stromal compartment is comprised of nonmalignant cells including immune cells, inflammatory cells, smooth muscle cells, pericytes, myofibroblasts, and vascular endothelial cells (2). Mobilization of these normal cells by malignant cells is critical to continuing tumor growth and malignant progression (1, 2, 4, 5). However, the genes expressed by malignant cells that enable invasion by stromal cells have yet to be fully elucidated. Many human tumor xenograft models have enhanced take rates and growth in immunodeficient mice when tumor cells were implanted along with stromal cells, activated fibroblasts, and/or extracellular matrix materials (2, 4, 28, 29, 30, 31). Many stromal-derived growth factors are involved in tumor progression, including urokinase plasminogen activator, hepatocyte growth factor, transforming growth factor-β, and epidermal growth factor (2, 4, 6, 32, 33, 34).

Several studies have used SAGE to identify genes involved in malignant invasion, and have used SAGE libraries of colorectal and pancreatic cancer to compared gene expression in normal and cancer cells (35, 36, 37, 38, 39). Ryu et al. (37) applied the biocomputational tools, Cluster and Treeview, for hierarchical clustering to identify genes associated with invasion in 10 SAGE libraries including 2 samples derived from normal colon mucosa, 2 primary colorectal cancers, 2 colon cancer cell lines, 2 pancreatic cancer samples, and 2 pancreatic cancer cell lines. The result was 90 SAGE tags associated with invasion that mapped to 74 known genes. In situ hybridization was used to characterize 12 of the genes in tissue sections of human pancreatic carcinoma (35). Eight genes were expressed within the stromal and/or angioendothelial cells, 4 of the genes were expressed by the stromal cells immediately adjacent to the invasive neoplastic epithelium, and 4 genes were expressed by the invasive neoplastic epithelium. Iacobuzio-Donahue et al. (35) elucidated genes associated with invasion by infiltrating breast carcinoma. SAGE tags from 11 libraries including 2 normal mammary samples, 2 ductal carcinoma in situ samples, 2 invasive breast cancers, 2 lymph node metastases, and 3 breast cancer cell lines were analyzed (40). One hundred and three SAGE tags associated with invasive breast carcinomas corresponding to 68 known genes were identified. Six genes were investigated by in situ hybridization and were found in five different regions of the tumor corresponding to different cell types.

The current study used human tumor cell lines from several major tumor types in cell culture with human dermal fibroblasts and HMVECs. An earlier study found that HMVECs in coculture with the tumor cells do not invade tumor cell clusters from any tumor cell lines, whereas myofibroblasts in coculture with the tumor cells do invade tumor cell clusters, and 1:1 mixtures of HMVEC and myofibroblasts allow both cell types to invade some of the tumor cell clusters (19). Among the well-established human tumor cell lines used in this study, the MDA-MB-231 cells have been recognized as invasive/metastatic in vivo when grown in nude mice (41, 42, 43, 44), whereas the prostate cancer cell lines such as LNCaP have been recognized as less invasive/metastatic (45, 46). The SAGE libraries analyzed in this study were from tumor cell lines on the assumption that genes expressed by the tumor cell lines in monolayer will allow the identification of genes important to malignant invasion. The initial GeneSpring clustering did not group the tumor cell lines by differential behavior in the experimental system, in agreement with the notion that only a small percentage of expressed genes is involved in malignant invasion (20). Refinement of the clustering analysis allowed the identification of several genes that may be associated with a more invasive malignant phenotype.

Several of the genes identified in this study have been found previously to be involved in aspects of tumor invasion and progression. BST2, also known as HM1.24antigen and EMP24, is a type II membrane protein found in the endoplasmic reticulum and cell surface, and expressed at high levels by multiple myeloma cells (46). The promoter region of the BST2/HM1.24gene has a tandem repeat of three cis- elements for a transcription factor, signal transducers and activators of transcription 3, which mediates interleukin 6 response gene expression (47). Interleukin 6 is a differentiation factor for B cells and a paracrine growth factor for multiple myeloma cells. Vascular endothelial growth factor triggers the increase in the interleukin 6 secretion level by the marrow stromal cells and, in turn, enhances the paracrine interactions between myeloma and marrow stromal cells (48). The expression of BST2/HM1.24 gene may be regulated by the activation of signal transducers and activators of transcription 3 in response to the vascular endothelial growth factor secretion by the myeloma cells. A monoclonal antibody against HM1.24 can induce antibody-dependent cellular cytotoxicity of multiple myeloma cells in vitro and in vivo (49, 50). A humanized monoclonal antibody to HM1.24 is currently in clinical trial as an immunotherapy to treat patients with multiple myeloma (48, 50, 51, 52, 53, 54, 55). BST2/HM1.24 was selectively expressed by a wide variety of solid tumors in the current study as well as by 100% of the four lymphoma SAGE libraries studied.

Stathmin-like 3/SCLIP belongs to the stathmin/oncoprotein 18 family of microtubule-destabilizing phosphoproteins (56, 57, 58, 59, 60, 61, 62). It has 70% identity with SCG10 protein, and is involved in signal transduction and regulation of microtubule dynamics (59). Stathmin-like 3/SCLIP is expressed mainly in neural structures, and is found at comparable levels in neonatal and adult rat brain, suggesting a potential role not only in the acquisition, but also in the expression of differentiated neuronal functions (61). Although under normal circumstances stathmin-like 3/SCLIP is expressed mainly in the nervous system, it has been found in brain tumors and mammary gland tumors. Many proteins have been found to play important roles both in neurite and angiogenesis regulation, such as vascular endothelial growth factor (63), basic fibroblast growth factor (64, 65), pigment epithelium-derived factor (66), and thrombospondin-1 (67). Analysis of SAGE libraries found expression of stathmin-like 3/SCLIP in 100% of brain and liver cancers, 75% of pancreatic cancers, 50% of lung and ovarian cancers, 30% of breast and colon cancers and melanomas, and 25% of prostate cancers.

TNFRSF5/CD40 is primarily involved in humoral and cell-mediated immune response (68). TNFRSF5 antibodies reduced atherosclerotic lesion size in hyperlipidemic mice, presumably by reducing macrophage and T-lymphocyte recruitment (69). TNFRSF5/CD40 was found highly expressed on cervical carcinomas (70). Activation of CD40 induced matrix metalloproteinase 9 in the cervical carcinoma cells, thus facilitating invasion by the tumor cells (70). In other experimental systems CD40 has been associated with induction of matrix metalloproteinases (71, 72), cell-adhesion molecules (73), invasion through collagen gels (74, 75), and tissue factor expression (76). Analysis of SAGE libraries found the most prominent expression of TNFRS5/CD40 was in the normal ovary, normal kidney, and dendritic cells. Among the malignant cell and tissue libraries, expression of TNFRS5/CD40 was highest in lymphomas.

HGS or hepatocyte growth factor-regulated tyrosine kinase substrate is a FYVE finger protein phosphorylated on hepatocyte growth factor stimulation, and is associated with cell proliferation, motility, and spreading (77, 78, 79, 80). Hepatocyte growth factor and its receptor c-met have been implicated in promoting tumor progression (81, 82) and angiogenesis (83). This pathway is the target of many preclinical and clinical therapies (84). HGS/HRS is involved in transforming growth factor β1 signaling and in influencing transforming growth factor β1-mediated suppression of human T cells (85). HGS/HRS associates physically with transforming growth factor-β-activated kinase 1 and p21-activiated kinase 1, and, thus, may be involved in c-fos induction (80). HGS/HRS binds to the tumor suppressor protein, scwannomin or merlin, a protein commonly mutated in schwannomas or meningiomas, and was found to inhibit rat schwannoma cell proliferation and migration, as well as cause abnormal cell spreading (86).

In summary, the combination of the stromal invasion assay and SAGE analysis appears a useful tool for identifying genes involved in tumor-stromal interactions. Although the cell culture assay and initial SAGE analysis was confined to eight human tumor cell lines, genes identified through the analysis demonstrated that the model reflected the gene expression in the tumor tissues when queries of specific gene expression was expanded to 150 SAGE libraries. SAGE is a systematic and unbiased method with no presumption of specific genes that can identify known as well as novel genes that are involved in the tumor-stromal interactions (20, 87). The genes identified through this analysis may be useful as therapeutic and/or diagnostic targets.