3.1 In silico est profiling strategies

For the modern biologist, there are numerous computational strategies that can be employed to assay gene expression. Many of these are based on utilising collections of expressed sequence tags (ESTs), unique segments of cDNA with base sequences identical to at least part of the coding region of a gene [3]. Because a large number of ESTs from diverse organ- and disease-derived cDNA libraries are being deposited in different databases, EST libraries are therefore an ideal source for expression profiling since EST clone frequency is in principle, proportional to the corresponding gene's expression level in a given tissue [16, 17]. This article reviews the many open-source online tools that have been developed to aid the handling, analysis and exchange of gene expression data in the public forum.

The aforementioned EST and SAGE data collections represent virtual goldmines of information for the modern biologist. Furthermore, these libraries are excellent starting points for disease-related gene discovery. For example, the EST database (dbEST) currently contains > 28 million public entries. Nevertheless, in any expression profiling experiment, be it in vitro or in silico, appropriate considerations need to be taken into account. These include the quantity and quality of RNA, where increasing these factors will invariably increase the yield of reliable and comprehensive experimental results. Concern over these issues is reduced by performing computational expression profiling and further careful in silico analysis can significantly reduce the amount of lab work required. Another caution worth considering is the source tissue. Many of these tools allow the user to select micro-dissected tissue as apposed to bulk tissue, therefore making the gene expression profile generated more specific. It is apparent from the amount of genomic information assembled in databases such as Unigene, that efficient tools are needed to mine these collections in search of meaningful information. To exploit this large amount of information, computer algorithms have been developed for the discovery of both novel genes [18] and genes with limited tissue distribution and/or disease-specific expression [19].

3.2 Sagemap

One of these algorithms, SAGEmap, is an online tool, specifically designed to interpret SAGE data [18]. SAGE data from any source may be submitted to this repository and SAGE data from a wide variety of sources may therefore be studied. SAGE data from both bulk tissues and cell lines from various species are collected in SAGEmap. Data is available on all SAGE data, including tissue type, depositor and any treatment the tissue has undergone. Comparisons of individual SAGEmap tag libraries can be performed to provide a list of differentially expressed tags in specific tissue libraries. SAGEmap is a user-friendly tool maintained by the NCBI and undergoes frequent updating. However, one weakness associated with SAGEmap is that tags with a count of one are excluded due to single-pass sequencing associated with tag production. This has little effect on tags with high abundance (i.e. more than one count) but can result in the loss of tags with counts of less than one. Nevertheless, such analysis involving tags with high counts enables differences of statistical significance to be easily identified.

3.3 X profiler

The cDNA XProfiler is another tool that compares gene expression between two pools of libraries, where each pool can be a single library or a grouping of several libraries [21]. For example, a user may compare diseased lung tissue with healthy lung tissue or furthermore a user may compare two different types of diseased lung tissue. For a gene to be "present" in a library pool, there must be at least one EST sequence found in the UniGene cluster for that gene. XProfiler lists all the genes found in each pool and categorises them as unique or non-unique. XProfiler further classifies them into known or unknown genes. The results are finally tabulated to show how these genes are distributed between both pools (Figure 2). A typical X-profiler analysis proceeds as follows;

Firstly the user selects tissue types (organ of origin) for each pool. The user then decides what tissue preparation methodology they want to include for each pool (e.g. cell line, bulk tissue, microdissected etc.). The next step is to select the library protocol that is of interest (e.g. normalised libraries, subtracted libraries etc.). The user then decides what tissue histology is to be included in each pool (e.g. normal tissue in one pool and cancer tissue in another) and finally pools are named and the query is submitted.

X-Profiler populates the pools with libraries that correspond to the user requirements. Pools can then be reviewed and modified if necessary. Following review the pools are submitted for comparison.

The results page contains three sets of information:

Gene Expression alterations are classified as unique or non-unique according to the pools of interest. The possible outputs for unique genes include.

First Pool Gene is only found in first pool.

Second Pool Gene is only found in second pool.

First Pool or Second Pool The total number of genes in both pools. These genes are found in either pools or maybe both, but not in any other library.

First Pool and Second Pool Genes found in both pools, but not in any other library.

In this way it is possible to identify all those genes whose expression is significantly changed between both pools (e.g. normal versus cancerous bulk breast tissue as separate pools therefore identifying genes that are altered in breast cancer). Figure 2 shows results obtained from this example comparison.

3.4 Digital gene expression displayer

Digital Gene Expression Displayer (DGED) is a tool for the comparison of gene expression between two pools of libraries. It can be used to compare either cDNA libraries or SAGE tag libraries. In contrast to XProfiler, it treats the presence of a gene in a library pool as a matter of degree. It compares the amount of a gene in one pool with the amount of the same gene in another pool. This comparison is reduced to two numbers: the sequence odds ratio and measure of significance. The formula used in DGED to calculate the sequence odds ratio between two pools A and B is; (Sequences in A/Total Sequences in pool A)/(Sequences in B/Total Sequences in pool B). DGED results are ordered by this odds ratio, with all cases of "NaN" (not a number) topping the list. NaN occurs when the denominator of the equation is 0, i.e., there are no sequences of a gene in pool B. An advantage cDNA DGED has many over other in silico gene expression techniques is that the user may select microdissected tissue source over bulk tissue thus giving a more specific gene expression output [22]. Another strength of DGED is that unlike the cDNA xProfiler, which lists every gene (even if an EST is seen only once in a pool) in both groups, the DGED finds only the statistically significant differences, based on the sequence odds ratio and a Bayesian test.

3.5 Digital differential display

Digital Differential Display (DDD) is a powerful web-based bioinformatic tool for the identification of differential gene expression. DDD uses the EST profiles of normal and disease cDNA libraries represented in the NCBI UniGene database. DDD compares the number of assignments of ESTs from different libraries, or pools of libraries to a specific UniGene cluster [1, 23, 24]. Fishers' exact test is used to restrict the output to statistically significant differences (P ≤ 0.05). It is therefore straightforward for users to then omit non-significant results from subsequent analysis. The output from DDD provides a numerical value denoting the fraction of sequences from each pool that maps to a specific cluster [25]. An example of a DDD experiment proceeds as follows;

3.5.1 Description of ddd experiment

In a typical DDD experiment the user must select which tissue libraries are to be assinged to each pool. The pools will then be compared. DDD compares the EST constituents of various tisue types, depending on which libraries are selected thereby determining the relative representation of each sequence in the libraries being compared. The DDD output is in the form of a web file that has links to Unigene clusters that correspond to the EST's that are differentially expressed between the two tissues (Figure 3).

DDD uses Fisher's extract test to restrict the output to statistically significant differences (P ≤ 0.05) therby determining the statistical significance of the number of times sequences from the selected libraries are assigned to a specific UniGene cluster. This is a statistical test for analyzing categorical data when the sample sizes are small. It has been argued that Fisher's exact test is too conservative, and is not appropriate to a setting where the total number of data values in the contingency table is not fixed [26]. The Bayesian methods used in xProfiler have a less conservative statistical threshold. The analysis is also restricted so that genes with over 1000 sequences in UniGene are included. A limitation of DDD is that these requirements place limitations on the capabilities of the analysis. Unless there are a large number of sequences in each pool, the frequencies of genes are generally not found to be statistically significant. Furthermore, the wide variety of tissue types, cell types, histology, and methods of generating the libraries can make it difficult to attribute significant differences to any one aspect of the libraries.

3.6 Digital extractor

A major limitation of in silico gene mining approaches is the cumbersome nature of the subsequent data analysis. The output from DDD is a list of Unigene clusters representing known genes, and sequences without homology to known genes that are significantly altered between selected tissue libraries. To expedite this strategy the data derived from the DDD comparisons can be processed using Digital Extractor. This application provides for high throughput processing of DDD output, by performing automated annotation of the output clusters. Digital Extractor can be used to both compile the profiles of known genes differentially expressed and also to annotate those clusters containing cDNAs without homology to known genes [27]. It utilises Contig Assembly Program-3 (CAP3) for assembly of EST clusters, Repeat Masker to mask repetitive elements and BLAST for gene identification [27].

3.7 How these methods compare

4.1 Gene expression in health and disease

Homeostasis in healthy tissue is dependent on the expression of genes that ensure cells have the machinery to deal with everyday events and furthermore ensure our well-being. However, unchecked variations in gene expression levels in a cell often lead to the initiation and progression of a disease process, such as cancer. It is therefore a propriety in modern biomedical research to determine and compare what genes are turned on and off in disease tissue and normal tissue. Using the approaches described herein, groups of genes that are characteristic of disease and may also be driving the disease process can be identified. Such genes may furthermore provide attractive targets for novel therapies in our efforts to overcome these debilitating diseases. Due to the extensive information obtained from genome sequencing, many of these techniques output ESTs of known and unknown genes. It is therefore dependent on the individual user whether known or novel genes take priority for further studies.

4.2 Cancer-associated genes

Cancer is a genetic disease. Expression profiling, as a powerful genomic tool, holds great promise in cancer molecular medicine and cancer research. This is because cancer is a complex polygeneic and multifactorial disease, resulting from successive changes in the genome of cells and from the accumulation of molecular alterations in both tumour and host cells [28]. Such genetic alterations effect regulatory pathways and cellular processes such as proliferation, differentiation, cell cycle, DNA repair and apoptosis and can also lead to genetic instability, tumourigenesis, malignancy, and an invasive and drug-resistant phenotype. Therefore, an understanding of the molecular behaviour of tumours would aid their molecular classification and also aid the decision-making regarding therapeutic approaches [29].

Computational methods of transcriptional profiling have been applied to further the understanding of all aspects of cancer biology. For example the identification of highly expressed genes may provide significant information thereby enhancing our understanding of tumourigenesis or serve as biomarkers or prognostic markers of malignancy [24, 25, 30]. DDD has been used to investigate gene expression in a wide variety of cancers including breast, colon, lung, ovarian, pancreatic and prostate cancers [1]. In a study by Scheurle et al [1], these cancers were found to share similar expression profiles, a concept that was proven using other laboratory techniques such as RT-PCR. In silico methods have identified the kallikrein genes, KLK6 and KLK10 to be overexpressed in colorectal, pancreatic and ovarian cancers [24, 31].

As apposed to simply identifying and compiling lists of genes, many studies have displayed genes identified in silico to be of functional importance as exemplified in two recent studies investigating gastric and colorectal cancer respectively [32, 33]. In these studies, genes identified as being differentially expressed using DDD were confirmed to have importance in key aspects of tumour cell biology such as cell proliferation and invasion. In both studies altered gene expression was confirmed with PCR using ex vivo cancer tissue in comparison with normal tissue. RNAi was used to knockdown gene expression, which resulted in decreased tumour cell proliferation and invasion. DDD has recently been used to identify genes with promoter similarities and that are therefore co-regulated in colorectal cancer [34]. Similarly, genes downregulated in gastric cancer were recently identified using DGED. In this study by Yanglin et al [35] KCNE2, a downregulated gene, was identified as a novel gastric cancer associated gene. Furthermore, the functional importance of KCNE2 was highlighted whereby overexpression resulted in growth inhibition. A recent study has demonstrated the tissue specific gene expression of various tumour types [22]. Dennis et al [22] employed SAGE and DDD to display strict differences in the expression patterns of different tumour types that could be used as markers of the various tumour types and for a better assessment of patient prognosis and optimal, tailored therapy. The study by Dennis et al [22] also identified novel potential tumour markers. Lipophilin B was identified to have expression restricted to breast and ovarian cancers while glutathione peroxidase 2 was specifically enhanced in colon and pancreatic cancers.

The use of SAGE to identify genes associated with the latter stages of cancer has recently been displayed [36]. In a study by Shen et al [36] the expression of advanced breast cancer was compared with that of benign tissue libraries and to identify fifty-three differentially expressed genes to be correlated with breast adenocarcinoma, a subset of which were successfully confirmed by RT-PCR. Likewise, genes associated with breast metastasis have been identified using SAGE [37]. The role of estrogen in the progression of breast cancer has been elucidated using a SAGE approach [38]. SAGE analysis has recently been used to identify biomarkers of gastric cancer [39, 40]. In a study by Yasui et al [37] SAGE was employed to identify differentially expressed genes in three categories; in gastric cancer in comparison to normal gastric mucosa, in advanced gastric cancer in comparison with early stage disease and in lymph node metastasis in comparison with primary tumours. A custom 395-element cDNA microarray representing these genes was then fabricated for use in diagnostics. The study by Yasui et al [39] represents the translational application of in silico gene expression profiling.

A recent study applied in silico EST profiling techniques to characterise various cancers including liver cancer where Bcl-x2 was identified as a novel liver cancer-associated gene [41]. Similarly, a combination of SAGE and microaray analysis has been employed to identify biomarkers of bladder cancer [42]. In this study by Wang et al [42] UCA1 was identified as a specific and sensitive biomarker of disease, which could be detected in the urine of bladder cancer patients. Another study utilised microarray analysis in conjunction with the in silico analysis of CGAP EST libraries to identify loss of annexin A1 expression in breast cancer [43]. This study and those studies outlined herein exhibit the value of in silico strategies in discovering biomarkers with clinical relevance in cancer detection and disease classification.

4.3 Tissue-specific gene expression

Gene expression profiling techniques can be applied to the identification of tissue specific transcripts or clusters of transcripts. Successful identification of those genes with specific expression in specific tissue types will aid our understanding of diseases arising in these sites. The large number of publicly available cDNA libraries corresponding to different tissues can be exploited using techniques such as DDD to identify genes with tissue specific expression. This approach has recently been employed to identify transcripts with preferential expression in renal podocytes [44]. This was the first study to use DDD to predict cell type specific gene expression. The authors successfully identified a protein SLM2 to be specifically expressed in renal podocytes and upregulated in proteinuric glomerular diseases and furthermore to be involved in VEGF alternative splicing. The study by Cohen et al [44] displays the potential that a gene discovery technique such as DDD has to predict and further our understanding of cell type specific gene expression.