A Public Database for Gene Expression in Human Cancers

Introduction

The advent of new technologies has made analysis of overall gene expression patterns a practical reality (1, 2, 3) . A major goal of CGAP 4 has been the identification of genes that undergo alteration in expression during malignant transformation, in part by capitalizing on these recent technical advances. 5 However, the means to efficiently generate, publicly distribute, and analyze comprehensive gene expression data from multiple laboratories did not exist previously. SAGE was chosen to create a comprehensive quantitative expression database for CGAP, because this assay provides absolute transcript numbers in a digital format (3 , 4) and can be easily adapted to provide statistical comparisons of data from multiple laboratories. The basis of SAGE is to count expressed transcripts by sequencing a 14-bp ‘tag’ of the gene, which is normally sufficient for gene identification. The four bases of the 3′-most NlaIII restriction site plus the following ten variable bases define the transcript tag. Past SAGE experiments have revealed that a small, but likely important, fraction of a tissue’s expressed genes has sustained a substantial change in expression after malignant transformation (5 , 6) . Unfortunately, the large data sets from these studies are not easily accessible for most researchers, and the time and cost of generating new tumor specific data are generally prohibitive. The goal of this project was to create the means to disseminate to the research community, without restrictions, comprehensive gene expression data for the study of most common human cancers. To initiate this first public quantitative gene expression database, we cloned and sequenced SAGE tags from 25 different normal and tumor tissues and created a system for rapid transfer of the data to the SAGEmap web site. Custom tools were created for this web site to analyze the expression of a single gene in multiple tissues or to compare the full expression pattern between tissues. Furthermore, we demonstrate one use of this database by reporting differentially expressed genes found by comparing glioblastoma with normal brain tissue. This is one of many comparisons that may be performed on SAGEmap.

Materials and Methods

Tumor and Normal Samples.

The following samples were used to construct the first 25 SAGEmap libraries: two colorectal carcinoma primary tumors, four colorectal cancer cell lines, two normal colon epithelium samples, four ovarian carcinoma cell lines, one normal ovary surface epithelium, two ovarian tumors, one prostate cell line, one GBM primary tumor, one pool of five GBM primary tumors, one fibrillary astrocytoma, one glioblastoma cell line, two medulloblastoma cell lines, one normal human astrocyte culture, and two normal brain cortical white matter samples. One normal brain white matter library was made from a rapid autopsy sample with a postmortem delay of 3 h and 15 min and confirmed normal histology (provided by the Bryan Alzheimer Disease Research Center, Duke University). The other normal brain library was constructed from normal surrounding brain removed during the surgical approach to remove a seizure focus or a glioblastoma. For each glioblastoma sample a neuropathologist (R. E. M.) confirmed the appearance of a GBM (Grade IV astrocytoma), without necrosis, by examining adjacent stained sections from the sample margins. Immunohistochemistry and slide histology were performed on fixed tissue sections embedded in paraffin. HAM56, a mouse monoclonal antibody (DAKO Corp., Carpinteria, CA), was used to identify macrophage specific markers in primary GBM tissue sections.

SAGE.

Total RNA for use in the SAGE protocol was prepared from tissue or cell lines by ultracentrifugation over a cesium chloride gradient followed by mRNA selection by oligo(dT) cellulose (Life Technologies, Gaithersburg, MD). The SAGE libraries were constructed as described previously (3 , 7) . For the brain libraries, 24 colonies were screened from each library before large-scale sequencing to ensure that they had an insert containing SAGE ditags. Approximately 2500 bacterial colonies were randomly picked by a robot and arrayed in 384-well plates. Plasmids from each colony were purified and sequenced as described previously (8) . Transcript tags were extracted from the sequence files, using the SAGE software, v3.03. Tags matching linker sequence (approximately 4%) and duplicate ditags were excluded before analysis or posting on the web site. To estimate the total number of expressed genes in each library, the unique tags were matched to a list of presumed possible ‘real’ tags. 6 This list compromises only those tags that are unlikely to occur by random sequencing errors, in part, due to the number of times they were observed in a pool of 3.5 million transcripts. Statistical comparisons of transcript numbers and P values presented here were derived from the Monte-Carlo analysis in the SAGE software (5) . SAGE tags from colon libraries were a combination of newly generated tags for this project or from a previous study (5) .

Northern Blotting and RT-PCR.

For northern analysis, total RNA was isolated by ultracentrifugation over a cesium chloride gradient from five normal brain samples, seven glioblastoma primary tumors, three cell lines, and six xenografts. Radioactive labeled PCR probes were made from the following genes (with GenBank accession number in parentheses):

• (a) α-glucosidase (X87237);

• (b) p1-cdc46 (X74795);

• (c) cartilage-39 glycoprotein (M80927);

• (d) nmb (X76534);

• (e) nicotinamide N-methyltransferase (U08021);

• (f) SPARC (J03040);

• (g) insulin-like growth factor binding protein 2 (M35410);

• (h) CAPL (M80563);

• (i) CSRP2 (U57646);

• (j) Eph-family protein (D83492);

• (k) SH3GL2 (X99657);

• (l) hCDCrel-1 (U74628); and

• (m) EST (AI214960).

Total RNA was electophoresed on an agarose gel and blotted overnight. Bands were normalized to β-actin and compared using autoradiography.

To confirm expression changes by PCR, total RNA was converted to random-primed cDNA using reverse-transcriptase for a panel of three normal brain samples, six glioblastoma primary tumors, two cell lines, and two xenografts. PCR of each cDNA was performed for 20, 25, 28, and 32 cycles using primers specific for β-actin, lipocortin (GenBank accession X05908), three ESTs (GenBank accessions AI302970, AI022985, and AA903288), and dynamin (GenBank accession L07807). The PCR reactions were normalized to the β-actin levels, and the products were compared on gel electrophoresis by band intensity to confirm large changes in gene expression between tumor and normal.

SAGEmap Informatics.

Tag-to-gene mapping was accomplished by first orienting GenBank sequences using polyadenylation signal and tail and placing the sequences into five “confidence” classes. In each of these confidence classes, tags were extracted from the ten-base tag directly 3′-adjacent to the 3′-most NlaIII site, and then linked to a UniGene cluster identifier, based on the sequence’s UniGene cluster assignment. This full mapping was used to provide the tag-to-gene and gene-to-tag information on the web site. In addition, a ‘reliable’ tag-to-gene mapping was designed to reduce the effects of sequencing errors in EST GenBank entries. To correct for these errors, tag-gene pairings were first ordered by the frequency that a tag was observed in a particular cluster. Assuming the errors occur randomly and a sequencing error rate of 1% per base, 10 bases have roughly a 10% chance [1-(0.99)¹⁰] of having one or more errors. Therefore, the lowest 10% of the rank-ordered tag-gene pairings were eliminated. It was this mapping that is used to make tag-gene assignments for SAGE tag frequency on the web site. The above approach was written into an automatic algorithm, which is used to update SAGEmap weekly.

To avoid the random simulations required to assess statistical significance by the SAGE software (5) , which would be impractical for an interactive web site, SAGEmap uses a previously described Bayesian approach (9) . We have extended this method so that it can deal with unequal aggregate numbers of tags A and B from the two libraries being compared. For an mRNA species with concentrations y and z in these libraries, we study the quantity x = y ÷ (y + z), which we assume has a prior probability density f(x) over the interval zero to one. This function captures knowledge about the distribution of relative mRNA concentrations for the complete set of mRNA species. It is reasonable to assume that f(x) will peak at, and be symmetric about, 0.5 (5 , 10) , and, for simplicity, f(x) may be taken proportional to x^c(1 − x)^c. Setting c = 1 has been proposed (9) , but various data (5 , 10) suggest a greater concentration of x near 0.5, implying c = 3 is a more appropriate value. If a and b copies of a tag corresponding to a specific mRNA are sequenced from the two libraries, the posterior probability density for x (i.e., the density given these data) is proportional to g(x), given by the following equation (derivation omitted). One may calculate from g(x) the posterior probability that the concentration y exceeds z by any given factor. Specifically, y exceeds z by at least the factor F when x ≥ L, where L = F ÷ (F + 1). The desired probability is simply the proportion of g(x), viewed as a density, that falls greater than L. The posterior probability that this is the case is given by the following equation.

Results and Discussion

More than 1,000,000 transcript tags from 25 different tumor or normal tissue SAGE libraries were cloned, sequenced, and deposited in the SAGEmap database using the project organization shown in Fig. 1 ⇓ . Six of these libraries were designed to study glioblastoma pathogenesis. Some of the possible transcript number comparisons from these libraries are shown in Table 1 ⇓ . From these comparisons we sought one that might provide insight into the pattern of expression necessary for glioblastoma maintenance or formation. Glioblastoma is a deadly, highly malignant form of brain cancer thought to arise mostly from astrocytes present in cortical white matter (11) . Therefore, we chose for further analysis the expression comparison between surgically removed glioblastomas and normal brain white matter. Tag counts generated by sequencing the libraries were used to determine transcript expression levels and to find statistically significant differences between the expressed genes. This comparison yielded the expected distribution (5) of the 282,760 SAGE tags (Fig. 2) ⇓ and showed that 1.0% (471) of the unique expressed genes had more than a 5-fold expression change between the tissues, and a P < 0.001.

By matching tag sequences to the predicted tag from GenBank human transcript entries, we identified well-characterized named GenBank entries to 160, or 34%, of the differentially expressed tags. Forty-three percent of these tags matched one or more ESTs. For the remaining 23%, there was no obvious matching GenBank entry, although one could retrieve the gene experimentally by using a PCR-based library screen (10) or by searching other sequence databases for a match. This latter group suggests that there are a significant number of expressed genes in these tissues without a GenBank sequence entry.

Proper tag-to-gene mapping was obvious when the tag identified an accurate full-length sequence, such as most of the named GenBank entries. However, sequencing errors in ESTs can create ambiguity in the mapping process. To enhance tag-to-gene mapping for ESTs, an algorithm based on UniGene 7 was constructed to ascertain mapping reliability. UniGene has sequence similarity clusters built from overlapping sequences and may have different tag sequences within the same cluster. However, SAGEmap takes into account the number of times a particular tag sequence is observed within each cluster, the orientation of the sequence, and the presence of a polyadenylation signal and tail. Thus, the likelihood that a particular tag is real and is not generated by sequencing errors is assessed and used for proper mapping.

We also constructed a tool for on-line statistical comparisons of tag populations for SAGEmap (Fig. 3A) ⇓ . This function allows comparisons between any possible combination of libraries and provides a confidence level for each differentially expressed tag. Desired fold differences can be specified to select the desired tags and libraries for comparison to suit the particular needs of an investigator. A ‘virtual northern’ function is available that calculates the fractional representation of a particular gene in all of the posted libraries (Fig. 3B) ⇓ . In addition, tag frequencies and mapping information are downloadable for local analysis.

Our next step was to validate and analyze a sample of individual genes with altered expression in our glioblastoma versus normal brain comparison. At least seven of the overexpressed genes were previously reported as such (Table 2 ⇓ , bottom). Eighteen other differentially expressed genes were tested by Northern blot, or when the gene of interest had low expression levels, by RT-PCR. Differential expression was confirmed in the appropriate direction using the original tumors, suggesting that tag-to-gene mapping and the observed tag differences were correct. However, altered expression in the original samples did not normally predict altered expression, by Northern analysis, in more than one-third of other glioblastomas (not shown). This likely reflects the molecular heterogeneity of these tumors (12) . Nevertheless, gene alteration in only a small fraction of cases may implicate a pathway involved in tumor pathogenesis.

Functional classification of these glioblastomas’ differentially expressed genes was used to provide insight into the pattern of altered expression (Table 2) ⇓ . Although comparing tumor with normal tissues produces an average gene expression difference, the observed changes often reflected our present understanding of tumor biology. For example, there is an expected loss of gene expression related to neurotransmission. The expression comparison was consistent with bone or cartilage metaplasia, which can occasionally be observed by histology in GBMs (11 , 13 , 14) . A group of genes with altered expression were related to macrophage function, which was supported by probing the original tumors with a macrophage-specific antibody. This experiment revealed positive-staining regions adjacent to necrotic regions of the tumor but no staining in normal brain areas (not shown). Also observed in the tumor sections was evidence of angiogenesis, a well-documented feature of glioblastoma (15) . This phenomenon was reflected by the 5-fold overexpression of VEGF as determined by SAGE. It is likely that in the 471 differentially expressed genes, there will be others involved in the ability of the tumor to sustain adequate blood supply but not revealed by our functional classification. Experiments aimed at ferreting out these genes from the candidate pool may provide useful therapeutic targets.

SAGEmap is the first completely public database for quantitative gene expression comparisons and provides a central repository for SAGE data. Most major tumor types and normal tissues are planned for representation on this database as well as comparisons designed to provide insight into the major genetic pathways involved in malignant transformation (7 , 16, 17, 18) . Another practical advantage of this data is that they provide candidate tumor-specific genes for immune-based cancer therapy (19 , 20) . SAGEmap now makes some large-scale expression data and comparisons quickly accessible to any researcher with Internet access. These data demonstrate one aspect of the utility of a gene expression database that provides a prototype for future reporting and dissemination of quantitative gene expression data. It is hoped that the virtual experiments possible using this database will accelerate certain aspects of cancer research.