Our approach was to exploit the entire collection of human EST sequences from dbEST 34 to obtain transcripts from different type of cells/tissues/organs. The assumption was that the activities of the genes can be represented by their transcripts, and also reflected by the number of representing ESTs in the NCBI dbEST database, given that a large number of mRNAs (cDNAs) were sequenced. Pertinent sequences from different sources were matched to genes and tallied together. Through the annotation of each EST record, we obtained the tissue type and condition type (normal or cancer) from which it was derived. With the information, we then had the entire gene transcription profile for all the tissues and conditions. Next, cross examining data of other sources including microarray data, secretory and membrane associations as well as analyzing protein associations with STRING 36 allowed us to narrow down the list of candidate genes. The process is illustrated in Figure 1.

The NCBI Reference Sequences (RefSeq Release 38, November 11, 2009) 37 were downloaded from its ftp site 38 . Homo sapiens RefSeq records were selected and subjected to repeat masking via RepeatMasker 39 .

Human EST data (Released on December 11, 2009) and their cDNA library information were downloaded from NCBI dbEST database 34 and CGAP 40 . Program in Python language was written to mark for discard the unsuitable libraries when the keywords such as "enrichment", "subtract", "pcr", and "normalized" were found in the DESCR, UNIQUE_PROTOCOL, or KEYWORDS fields of the library information. An arbitrary cutoff of > 400 was chosen to the highly unrepresentative libraries (approximately 7,000 libraries constituting approximately 650,000 ESTs were discarded as a consequence). To curb from incorrect inclusions or exclusions, we finalized the process with manual curation. Libraries made from mixed tissues or cell lines were also discarded. The final libraries from CGAP were manually classified into 48 different tissue types and two different conditions, normal and cancer.

The BLAT alignment tool was used to align ESTs to RefSeqs as a mean to assign ESTs to genes 41 . The criteria of having an identity of 95% or above and the minimum length of 100 nucleotides were set for a match. The RefSeq match with the highest identity was assigned for the EST. If two RefSeq matches shared exactly the same identity, the program chose the first encountered.

The procedure attributes each transcript represented by RefSeq its expression profile across different tissue and condition types based on EST assignment counts. Each EST has its corresponding tissue type and condition type classification, based on its source clone library. For example, a transcript with an aligned EST from a lung cancer clone library is one expression count each in tissue type lung and condition type cancer. This way, after all ESTs were counted, each transcript has a profile of expression across various libraries and conditions. Expressions from different transcript variants of the same gene were pooled to obtain a single gene expression. The raw counts were thus made into transcription profile for each gene for further statistical analysis.

Cochran-Mantel-Haenszel statistics (CMH) was applied to evaluate cancer differential expression of each gene. To evaluate each gene, other genes were pooled as "other genes" to create a 2 × 2 × k table consisting of data from tissue-condition cross, where k was the number of tissues × 2 (two conditions). A contrived example of 2 × 2 × k table where k is 2 is shown in Table 1. Gene A is the gene under study while other genes are pooled together as "other genes". Only Tissue I and Tissue II columns are calculated in CMH. The pooled ones are not part of the analysis. Akin to Fisher's exact test, the test assumes that "other genes" should consist mostly of genes not differentially expressed between normal or cancer conditions. Or, some of them are DEGs for one condition, but they are at least partly canceled out by DEGs for the other. In any case, the imbalances of cancer counts to normal counts in the second row is regarded as owning to sample bias and it serves as a metric against which Gene A is measured. By continuously isolate values for gene currently under study while pooling all other genes to the second row, an odds ratio and a confidence interval is calculated for each gene. Genes with a p-value < 0.05 and an lower bound of confidence interval of odds ratio ≧ 1.65 are selected for further analyses.

Human U133 Plus 2.0 GeneChip array CEL data were downloaded from Gene Expression Omnibus (GEO) 42 . When computing power allows, the data were processed with AffyPLM 43 using its three-step procedure of processing background signals with GCRMA, normalizing signals with quantile normalization, and summarize probe signals with medium polish. For large experimental datasets that were computationally infeasible for us, we used justRMA from the Affy package 44 . For experimental dataset without raw CEL data, we obtained the pre-processed matrix files via GEOQuery 45 . Regardless of the source of array signal processing, we analyzed the genes for differential expression with Limma 46 . Differentially expressed gene candidates with p-value < 0.05 and logFC > 1.0 were selected and crossed with genes from EST profiling with statistical evaluation. For each array, the significant genes were crossed with our EST profiling results. The union of these intersecting genes was selected for further evaluation.

To identify our differentially expressed genes with secretory annotation, a list of 3,975 proteins with secretory annotation originated from the conglomeration of data from Uniprot (1,632 unique proteins) 47 , Human Plasma Proteome Organization (HUPO) (889 proteins), and Secreted Protein Database (SPD) (4,142 proteins) 48 . This list was matched against DEGs to give them secretory annotation.

Membrane protein annotations were gathered from five sources - TOPDB (283 proteins) 49 , LOCATE (2629 proteins) 50 , PDB_TM (41 proteins) 51 52 , OPM (107 proteins) 53 , and MPDB (23 proteins) 54 - to generate a unique list of 2,767 membrane proteins. Any DEGs on this list would confer it a membrane annotation.

TissueScan™ Cancer Survey Panel 96-I qPCR array panel (Origene Technologies, Rockville, MD) containing the cDNAs of 3 normal and 9 cancer tissues each from 8 organs (breast, colon, kidney, liver, lung, ovarian, prostate, and thyroid) was used to examine the expression profiles of selected cancer differentially expressed gene candidates. Real-time qPCR analyses with the Taqman^® Gene Expression Assay kits (Applied Biosystems, Foster City, CA) and FAM- and VIC-labeled target genes and HPRT1 internal control primers, respectively, were performed according to the manufacturer's suggested procedure on an Applied Biosystems Prism 7500 system. Relative specific gene expression was quantified by normalization against the HPRT1 with the ΔCT method. Gene expression changes were quantified as 2 ^{- (CT gene - CT control)}.

The basic steps of our analysis are illustrated in Figure 1. A total of 8,296,089 human EST sequences (Dec. 11, 2009 release) were downloaded from the NCBI. Despite the size of the data, not all ESTs are relevant for our gene expression analysis. After screening the 8,907 EST libraries as described in the methods section above, 8,447 unsuitable libraries, the preparation of which involved PCR amplification, normalization, subtraction, etc. or originated from cell lines, were discarded. The remaining 460 libraries consisted of 2,386,536 EST sequences representing approximately a third of all the downloaded human ESTs.

After BLAT alignment of the 2,386,536 ESTs to 44,513 gene transcripts from RefSeqs, approximately 1,644,960 (68.92%) ESTs with at least 100 nucleotides matched to RefSeqs were detected. An examination of the sources of the matched ESTs indicated that the representativeness of each tissue is skewed and that the brain is the most represented out of all tissues. Among the 48 different tissues, brain ESTs constituted 26% of all matched ESTs, uterus (6.40%) ranked second, followed by testis (5.91%), placenta (4.33%), pancreas (3.99%), muscle (3.88%), liver (3.51%), kidney (3.52) and others each below 3% (see Additional file 1). Similarly, condition type (normal and cancer) representation was also skewed. Normal tissue type had 1,251,883 ESTs combined, and cancer tissue had 393,077 ESTs in the ratio of roughly 3 to 1. Originally before filtering out those from the cell lines, there were more cancer ESTs and the ratio of normal ESTs to cancer ESTs was roughly 1 to 3. This showed how much more rigorous our filtering was. Unfortunately, this also meant we had a much smaller dataset to work with.

The unequal distribution of the 1,644,960 matched ESTs in different tissue types caused some tissue types to be ill-represented. For example, the number of brain EST hits dominated over other tissue types. On the other hand, spinal cord had the least count with 430 EST hits. The latter had little value for our application. Therefore, we only took a tissue type into consideration when its total EST hit count was above the cut-off of 20,000. Considering that the human genome has approximately 22,000 genes, the cut-off still did not allow "deep" probe into gene expression. Nevertheless, the method we employed did not attempt to identify specific gene expression in one particular tissue; therefore, the problem was mitigated.

We also categorized ESTs according to their clone library classification, to either be from normal or from cancer. Sometimes a certain tissue-condition type was so under-represented that the information was not trustworthy. For example, adipose had 10,362 normal hits but only 440 cancer hits, and heart tissue had 22,179 normal hits but no cancer hits. For these cases, data was kept throughout the analysis. But these data did not make contribution to our analysis.

Since our EST assignments were made to transcripts represented by RefSeq sequences, when the entire assignment procedure was done, each transcript variant had its expression profile across all tissue-condition types. Due to the lack of enough ESTs data, differentiating between different splicing variants of the same gene was not feasible. We had to pool expression from different splicing variants into a single expression profile representing the gene.

Due to the small sample size (EST counts), it was only realistic to evaluate gene expression based on all ESTs of all tissues. However, tissue type was a confounder. If all counts for each gene were pooled as "normal" or "cancer" regardless of the tissue of origin, the count would be incorrect. To solve both the sample size and the tissue confounder problems, Cochran-Mantel-Haenszel statistical method was employed to identify genes with differential expression as described in the method. We used the arbitrary cut-offs of p-value < 0.05 and odds ratio ≧ 1.65 to obtain a primary set of candidates. As a result, a total of 723 cancer differentially expressed gene candidates were selected. The 1.65 cut-off is chosen based on a good coverage to a list of well-known biomarkers or genes known to associate with cancer (Table 2).

To show that this list of 723 genes was enriched for cancer and thus obtains credibility for our methodology, we looked for cancer related pathways associated with them in GeneGO 55 pathways, which covered 650 signaling and metabolic networks (Figure 2A). Among the 10 most significantly matched pathways, several are cancer related - Pathways number 1, 3, and 4 involve immune response; number 2 and 5 involve cytoskeleton remodeling; number 6 is transition and termination of DNA replication; and number 8 and number 9 are adhesion related. In addition, the result of GeneGo disease enrichment analysis (Figure 2B) indicates our set of genes as neoplasm enriched: seven out of the 10 most associated diseases are related to cancer. The disease ranks the highest is neoplasms, followed by neoplasm by site, and digestive systems neoplasm. This list reveals that our 723 DEGs covers general neoplasm related functions, and not specific to any particular neoplasm, as digestive, urogenital and breast are all covered.

To narrow down this list of biomarkers, we crossed examined the expression profiles of the candidates with the differentially expressed genes in 6 microarray experiments, i.e. two each of ovary and uterus, and one each of pancreas and colon (Table 3). These tissue types were selected based on the following reasons. We noticed that many of our candidate genes had the most expression in ovary tissue (after normalization). The other concern was the number of ESTs. Since our candidate genes were derived from EST sampling of various tissue types, they were influenced more heavily by tissue types with more EST representation due to deeper sampling from them. Therefore, the rest of the tissue types were selected based on their representativeness. Of the 723 DEGs, 235 candidates were also found to be differentially expressed genes in our microarray analysis.

Since membrane and secretory proteins could be potential therapeutic target or serum biomarkers, the subcellular location of the 235 DEGs were examined against the secretory and membrane protein lists consolidated from public databases. Among these, 96 DEGs were putative membrane or secretory proteins - 57 had only secretory annotation, 27 had only membrane annotation and 12 had both.

To further examine whether the 96 membrane/secretory DEGs identified in our EST database mining had enriched cancer-related genes, we searched the literatures for known associations with cancers. In additions, they were also analyzed with STRING for interactions, which are based on experimental evidence or prediction, such as conserved genomic neighborhood, gene fusion, co-occurrence across genomes, pathways, protein complex, co-regulation, or other literature sources such as co-mentioning. The network of the STRING interactions of the 96 DEGs together with the literature search results were plotted based on the combined STRING score with Cytoscape 56 (Figure 3). Approximately 68 proteins formed a big cluster of interacting proteins and a large proportion of the DEGs (88%) had published cancer association with clinical or non-clinical experimental supports. This demonstrates the value of our integration strategy since we had an ample of literature supports.

The 96 DEGs were selected out of their general cancer propensity without necessarily referring to any particular tissue type. However, we can still assess the general tissue distributions shown in Figure 4. A gene has a tissue representation if any EST from a clone library of the tissue type is matched to it. We can see that some genes are observed across many tissue types. A gene could be observed across a variety of tissue types if it is pan-tissue, and its expression measure is relatively abundant. Separately, Woolf's test for heterogeneity can also give hints to whether a gene is pan-cancer. Those that were found as significant in this test were considered having unequal representation in different genes; although whether they are pan-cancer require further evaluation.

Three cancer differentially expressed secreted protein gene candidates, COL3A1 (Collagen alpha-1(III) chain), DLG3 (Discs large homolog 3), and RNF43 (Ring finger protein 43), which had an odds ratio of 3.55, 7.97, and 4.03, respectively, and with limited or no clinical support were selected for real-time qPCR analysis using the Taqman^® Gene Expression Assay kits (Applied Biosystems, Foster City, CA) (Figure 5). With the HPRT1 as the reference, higher expressions of these genes were noticed in at least some of the cancer tissues. Apparently, the average relative expression levels of COL3A1 in breast, liver, thyroid cancer samples were higher than their normal counterparts. The average expression levels of DLG3 in breast, kidney, liver, lung, and ovarian cancers, and RNF43 in colon, liver, lung, ovarian, and prostate cancers were also found to be higher than their normal tissues. The expression of COL3A1 in approximately 5 of the liver cancers, DLG3 in 5 of the liver, 7 lung and 5 ovarian cancers, and RN43 in 7 of the colon, 8 ovarian and 5 prostate cancers seemed to have higher expressions than the normal tissues. In light of the limited sample size, the three candidates appear to have an overall higher expression in cancer tissues.

The meta-analytic nature of our study brought us the opportunities as well as challenges to study the digital signatures of various transcriptomes in a new perspective. Comparing to experimental methods that focus on a single tissue type or limited tissue types, our approach allows us to find genes inclined to express in cancer in a pan-tissue manner. An important challenge of our approach is to avoid Simpson's paradox which can occur in a meta-analysis study 32 . Simpson's paradox is where the association between two variables may show a correlation that is reversed in direction from what is observed from stratified subgroups. A contrived example is shown in Table 1 in which gene A appears to have a higher cancer expression when pooled, but it is in fact not so under individual stratified sub-tables. This may be somewhat of an extreme case, where directionality of the ratios actually differs between the sub-tables and the pooled table. However, the tissue confounder still introduces bias, large or small, that may throw our judgment off. In this study, we used CMH to analyze the data based on stratified sub-tables to avoid running into this paradox. One could also analyze only one tissue at a time for differential expression, but this means one has a smaller dataset to work with. CMH could avoid this problem since it uses EST counts from all tissues instead of analyzing just the normal and cancer propensity under each individual tissue type.

In our actual data, the odds ratio of the pooled table is also different from that of the stratified table. For example, the gene PTRF, a polymerase I and transcript release factor, has a pooled odds ratio of 0.40 and a CMH odds ratio of 0.16 calculated from stratified sub-tables. In this particular case, both odds ratios indicated an inclination toward a higher normal expression and are both statistically significant although at different degree (the pooled has a p-value of 5.269e-15 under a 2x2 χ² test 57 versus CMH's 7.16E-77). For the gene VCAN (versican), the pooled odds ratio is 1.86 and χ² test yields a significant p-value of 1.83e-4. However, CMH gives an insignificant result for this gene with p-value of 0.25. As an extreme case, GBP6 (guanylate binding protein family, member 6) has a pooled odds ratio of 6.69 and χ² test gives a p-value smaller than 2.2e-16 (approaching 0), whereas with CMH the odds ratio is 0.73, actually indicating a higher normal counts, although CMH p-value of 0.15 is insignificant. This indicates Simpson's paradox in action. Careful inspection showed that all cancer counts and most normal counts of GBP6 were contributed by the tongue tissue source. Out of a total of 50 cancer counts and 21 EST normal counts, tongue accounts for cancer and normal counts of 50 and 17, respectively. For this gene, the tongue cancer count 50 is not influential under a total of 29,479 cancer counts and 7,486 counts for the tongue. Thus pooling loses information in this respect and gives a false impression that its cancer expression is much higher when summing all cancer counts from all tissues. Stratifying by tissue type guards against this bias.

In the strictest application, the use of Cochran-Mantel-Haenszel method requires the odds ratios of the sub-tables be homogeneous. In our context, it means the ratio of gene expressions between cancer and normal tissues are probably the same among all tissue types under study and any observed variability is most likely due to sampling bias. Also, the calculated odds ratio would be the estimated common odds ratio across the tissue strata. In our case, however, not all genes had similar ratios under each tissue (based on Woolf's test for homogeneity available in Additional file 2 under the "Woolf" column label), and this was of course expected. In spite of this, we were interested in the overall expression patterns of the genes in cancer conditions. We were not interested in an estimate of common odds ratio across the strata, which often does not exist. We were interested in hypothesis testing - to give us leads to the genes that had higher cancer expression in general. In this regard, the test could be applied 58 59 . The CMH odds ratio is a weighted average of the odds ratio in each tissue classification and can give us a summary measure 60 , which we used to prioritize and followed up with subsequent biological analyses. In other word, an odds ratio in our data was merely a value that "average up" across all tissue types. From these ratios we were able to reveal the preferential cancer expressions, since the list covered a number of important known biomarkers, and enrichment of cancer-related genes were supported by knowledge-based GeneGO analyses and previous publications.

Another distinctive tactic we used is the selection of DEGs among the statistically significant genes (p-value < 0.05) base on lower bounds of the confidence interval of the odds ratio estimates. The popular approach to search for DEGs is to select genes base on p-value first, and then select the subset base on parameter estimators such as odds ratio or fold change values. The p-value criterion selects the statistically significant ones (those not likely to be the result of random fluctuation). The subsequent criterion is based on prior domain knowledge. However, among those with statistically significant p-values and similar parameter estimators, the ranges of the estimations can vary widely. Using our dataset as an example, the two genes TUBA1B and FAM60A both have odd values of 2.38 (Additional file 2). However, for TUBA1B, it is within the 95% confidence that its true odds ratio is between 2.26 and 2.50. Yet for FAM60A it is between 1.59 and 3.54. Based on our background knowledge and for future application, if we must select genes having odds ratios greater than 2.0, then using odds ratio as cutoff would not serve this purpose since it is quite possible that the real odds ratio (i.e., of the population) is below 2.0. Choosing genes based on their confidence intervals would be more precise, but this has not been much appreciated.

The multi-cancer approach compares genes that are overall differentially expressed among multiple cancer types comparing to their respective normal tissue types. Although many biomarker studies focus on gene differentially expressed in a particular tissue type, Wu et al. found 8 proteins in the conditioned media of 23 cell lines showing negative or weak tissue staining in the Human protein atlas, suggesting them to be potential pan-cancer markers 61 . Sahin et al., found that claudin-18 splice variant 2 had the ectopic activations in pancreatic, esophageal, ovarian, and lung tumors while its expression in normal tissue only occurred in differentiated epithelial cells of the gastric mucosa, confirmed by RT-PCR 62 . These studies suggested that relatively multi-cancer genes or multi-cancer splice variants exist. The three candidates COL3A1 (Collagen alpha-1(III) chain), DLG3 (Discs large homolog 3) (plasma membrane), and RNF43 (Ring finger protein 43) are putative secreted or plasma membrane proteins with the potential of developing serum diagnostic reagents. In reviewing the involvement of these genes with cancers in previous studies, hint for pan-cancer marker was surfaced as the expression of the extracellular matrix protein COL3A1 gene in brain cancer 63 and angiofibroma 64 was elevated. While secreted membrane bound RNF43 protein gene was known to be up-regulated in colorectal cancer 65 . Interestingly, upon the real-time qPCR analysis of three cancer differentially expressed secreted protein gene candidates, COL3A1, DLG3, and RNF43 identified in this study, higher cancer expression levels of these genes in multiple cancer types were verified. This does not only indicate the usefulness of our computational approach and filtering procedure but also encourages us to devote further resources for assessing the clinical usages of these three candidates.

Earlier in this discussion, we mentioned that naïve pooling of data may introduce bias and at worst may produce Simpson's paradox. We also mentioned that we have tackled this problem with CMH. Nevertheless, two other occasions of pooling actually took place. We pooled expression from different splicing variants from the same gene to make one gene expression. We also pooled different libraries of the same tissue into one tissue classification. In both of these cases, we may encounter expression bias, since different splicing variants and different tissue libraries (i.e., tissues from different patients) might have differences in expression patterns. This is an unfortunate limitation in this and similar studies, since dbEST data consists of many different sources, and given the relative lack of data after the very stringent criteria we have used in our library selection compare to previous studies (Most importantly the exclusion of ESTs from cell lines, PCR amplification, subtraction, and cDNA normalization protocols). We opted for pooling since we had comparatively limited number of sequences to work with (1,644,960 out of 8,296,089 downloaded - 18.03%). Nonetheless, future digital expression profiling can be made better with the RNA-Seq methodology that offers a greater depth of coverage than ESTs obtained from traditional cDNA sequencing. It gives a much larger sampling size that makes more realistic the differentiation among isoforms and also makes pooling of different libraries of the same tissue less necessary. As for discovery of pan-cancer genes or isoforms when studying multiple tissue types, similar idea as outlined in this study would be just as applicable.