For the list of the datasets used in this study, please see Additional file 1. Gene expression datasets were retrieved from the Oncomine 3.0 database, accessed in May 2008 28. We used 18 datasets in total: 11 datasets for the transition from normal prostate (NP) to nonmetastatic prostate cancer (nMPC) and 7 for the transition from localized (nMPC) to metastatic prostate cancer (MPC). To analyze the genes involved in prostate cancer initiation – transition from normal to prostate intraepithelial neoplasia (PIN) – we used data from a study by Tomlins et al. 29. Individual P values detected in each study were used to compute global P values using meta-analysis.

Combining the results of multiple independent studies allows more robust statistical conclusions than does the analysis of a single dataset. However, including a flawed study in such a meta-analysis could lead to a bias; therefore, the quality of the data must be carefully considered before including a given dataset in the analysis. To ensure reproducibility of the data, we assessed the overlap of significant genes in independent datasets, our rationale being that studies targeting the same phenotype should be expected to detect an overlapping set of differentially expressed genes. We therefore excluded studies wherein the overlap in up/up- or down/down-regulated genes was lower than would be expected in such pairwise comparisons.

We assessed the overlap between significant genes (significant genes were defined by the liberal P value of 0.05) that were up- or downregulated in two studies. The results of the analysis were presented using the 2 × 2 data format. For each cell in the 2 × 2 table, we computed the expected number of genes. If N_1,2 were the number of genes analyzed in both study 1 and study 2, P_u1 and P_u2 would be the proportion of genes upregulated in studies 1 and 2, correspondingly. Thus, the expected number of genes upregulated in both studies could be computed as a product of the corresponding proportions and the total number of genes analyzed in both studies. For example, the expected number of genes upregulated in both studies E_u/u may be computed as E_u/u = P_u1P_u2N_1,2. A typical example is shown in Table 1. The observed number of genes consistently (in terms of direction, up or down) differentially expressed (up/up or down/down) was much higher than the expected number (χ² = 6,581.2, df = 1, P value extremely low – essentially zero for this statistic). Gene expression data from two independent studies 30 and 31 were used in this analysis. The observed numbers of genes not consistent in their expression (up/down and down/up) were similar to that expected (χ² = 2.1, df = 1, P = 0.15), suggesting that genes upregulated in one study but downregulated in another were likely to be false positives.

If the ratio of the observed to the expected number of genes were to be close to 1 for all cells in a 2 × 2 table, it would suggest that at least one study was severely flawed and should be excluded from analysis. In a comparison of just two studies, it would be impossible to say which one should be excluded. However, if all but one among multiple (i.e., six) datasets showed a considerable overlap, we could conclude that the inconsistent study should be excluded. On the basis of this approach, we excluded two studies from the original list of datasets.

We computed the expected number of significant genes based on the strict null hypothesis, which assumes that all findings are technical or statistical artifacts and that a random set of genes is detected as significant in each individual study. The excess of the observed over the expected number of genes would then be due to the presence of true statistical positives. The proportion of true positives among all significant genes can be computed as P_t = (Nobs - Nexp)/Nobs. We estimated P_t across different numbers of independent studies used in meta-analysis. We found that combining six or more independent studies in a meta-analysis provided statistically robust detection of true positives. Figure 1 shows the dependence of the proportion of true positives on the number of studies used in meta-analysis, with the combination of four independent studies reducing the probability of false positives to lower than 0.05. If six studies were included, the probability of false positives would be lower than 0.01.

Note that our approach for the assessment of the proportion of true positives is different from the estimate of false discovery rate (FDR) that is based on the Bonferroni correction and computed as a product of the type one error and the total number of tests 32. The FDR approach takes into account the possibility that false significant findings could result from multiple testing, whereas an inflation of statistics resulting from technical artifacts or poor study design would be treated as a true signal. Our approach was more conservative, allowing for the exclusion of false positives resulting from technical artifacts and flaws in study design.

We used an extension of Stouffer's method 33 to estimate the overall significance based on the significance of the individual tests. This approach is based on estimating the standard normal deviation Z and is similar to the approach recently proposed by Ochsner et al. 1. Selection of the method of meta-analysis for our study was dictated by the available data: for the majority of the datasets, the raw gene-expression data were not available, while t tests and corresponding P values were easily accessible for individual probes.

Individual probability P is first converted into a Z score and the Z scores summed up across studies. This sum is divided by the square root of the number of tests (k). The k was the number of datasets where the specific gene was assessed: maximum 11 for the NP>nMPC and 7 for the nMPC>MPC transition. The sum of normal deviates is itself a normal deviate and can be backtransformed into an overall P; the probability level associated with the sum of Z yields an overall level of significance. The complete procedure takes the following steps: (1) P_i → Z_i; (2) Z_(overall) = ΣZ_i/; (3) Z_(overall) → P_(overall). The advantage of this method lies in the increased power of the overall comparison. If, for example, several tests consistently favored the research question but failed to reach the level of significance due to small sample size, the overall test would more easily become significant because of the pooled sample size being much larger than its components.

For each gene, we computed the overall global P values based on individual P values from independent studies. We then ranked all genes according to the P value from smallest to largest. We used the top-ranked genes to analyze their clustering by pathway, molecular, and cellular functions.

For functional annotation, we applied Ingenuity Pathways Analysis (Ingenuity Systems, http://www.ingenuity.com/) and the Database for Annotation, Visualization, and Integrated Discovery (DAVID) 34. We analyzed the data at three levels: that of (i) gene, (ii) pathway, and (iii) biologic function. Ingenuity and DAVID use similar approaches for functional annotation of candidate genes, looking at the distribution of the top-ranked genes by pathways and/or gene ontology categories, testing the null hypothesis that genes are randomly distributed across pathways and biologic functions. P values characterize the statistical evidence for the clustering of the genes by pathways or functional categories; the lower the P value, the stronger the statistical evidence that the top-ranking genes belong to a specific pathway or functional category. In this study, we used the classification of pathways according to the Kyoto Encyclopedia of Genes and Genomes (KEGG) 35. The KEGG provides a collection of manually drawn pathway maps representing current knowledge about the molecular interactions for metabolism, processing of genetic and environmental information, cellular processes, and human diseases 36.

To decide how many top-ranked genes to include in functional annotation analysis, we analyzed the dependence of the number of significant pathways and functional categories on the number of top genes (see section "How many genes to use for an assessment of functional clustering" in Additional file 2 for details of the analysis). On the basis of the results of the analysis, we decided to use the top 500 genes for functional annotation.

Because we analyzed many genes, it is possible that some significant findings could be the result of multiple testing. Figure 2 shows the distribution of P values for the genes tested in the transition phases from NP to nMPC (NP-nMPC) and from nMPC to MPC (nMPC-MPC). For both transitions, a significant excess of low P values was observed, suggesting the presence of true positives.

We found that several adhesion pathways, including focal adhesion, integrin signaling, and tight junction signaling, were significantly associated with prostate tumorigenesis. Therefore, we decided to look at the expression of cell adhesion genes in more detail. We stratified cell adhesion genes into five categories: cadherins, immunoglobulin-like adhesion genes (ICAMs), integrins, selectins, and tight junction genes. We found that integrins tended to be downregulated in prostate nMPC compared with NP (see Additional file 9). Of 24 integrins and the proteins directly binding them, 20 genes, or 83%, were downregulated, and only four were upregulated, compared with the expected 46% based on the overall analysis of significant genes at P < 0.05 (χ² = 11.4, df = 1, P = 0.0003). The excess of downregulated integrins was even more striking when only integrins themselves were considered: all except ITGAX were downregulated. Other types of cell adhesion genes (cadherins, ICAMs, selectins, and tight junction genes) also exhibited predominant (although less profound) downregulation (see Additional file 9).

Integrin ligands also showed a tendency to be downregulated. In the NP-nMPC transition, 30 of 38 differentially expressed integrin ligands were downregulated (Additional file 10). In the nMPC-MPC transition, the number of downregulated integrin ligands was similar to the number of upregulated integrin ligands: 32 and 28, respectively. The major sources of integrin ligands in prostate tissue are fibroblasts, although some investigators suggest that epithelial cells can also produce integrin ligands 37. We tried to distinguish between the two sources of integrin ligands by separately analyzing grossly dissected and laser-capture microdissected tumor cells. We found no collagen, fibrinogen, or laminin genes among the top 500 genes in the two studies that analyzed gene expression in individually dissected tumor cells 2938. In the studies that assessed gene expression in a mixture of tumor cells and fibroblasts, we found three differentially expressed collagen genes: COL15A1, COL4A6, and COL4A5. All three were significantly downregulated; the corresponding P values were 7.8E-12, 7.0E-11, and 1.1E-9. These results suggest that differences in expression of extracellular matrix proteins are likely due to differences in gene expression in fibroblasts.

We also looked specifically at the expression of integrins at the initial stage of prostate tumorigenesis – the transition from NP to PIN – using gene expression data from Tomlins et al. 29. Ten of 22 integrins analyzed in the study were differentially expressed between NP and PIN. Seven integrins (ITGA3, ITGA4, ITGA8, ITGB4, ITGA2, ITGAM, and ITGB3BP) were downregulated, with P values of 8.2E-5, 0.005, 0.008, 0.01, 0.03, 0.03, and 0.04, respectively. Four genes (ITGAV, ITGB5, ITGAL, and ITGB3BP) were upregulated in PIN compared with NP, with P values of 0.001, 0.0006, 0.009, and 0.02, respectively. When we looked at the expression of integrin ligands, we found that 21 genes (COL5A3, COL6A1, LAMB2, COL11A2, COL17A1, FN1, LAMA3, COL7A1, COL3A1, COL6A3, COL4A2, LAMB1, COL23A1, COL13A1, COL1A2, COL16A1, COL20A1, COL9A1, LAMA2, COL4A5, and COL15A1) were differentially expressed between NP and PIN. All ligands except LAMB1 were downregulated, suggesting that suppression of integrin ligands plays a role in the initiation of prostate cancer.

Additionally, we compared the expression of 69 cell-adhesion molecules: cadherins, integrins, selectins, ICAMs, and the genes involved in the formation of tight junctions between stromal and epithelial cells. We used the data from the study conducted by Tomlins et al. 29. The list of the cell-adhesion genes was based on the genes from two cell-adhesion databases 3940. Of 69 cell-adhesion genes, 63 did not show a difference in the gene expression level between stromal and epithelial cells. Of six differentially expressed genes, three – CDH1, ITGB5, and TJP3 – were downregulated and three – ICAM1, ITGA8, and ITGB5 – were upregulated in the stroma relative to the level in the epithelium.

We found that 77 or 15% of the top 500 genes involved in the NP-nMPC and nMPC-MPC transitions were the same. This overlap was much stronger than would be expected purely by chance. Indeed, 17,859 genes were assessed in both the NP-nMPC and nMPC-MPC analyses, making the proportion of the top 500 (500/17859) to be 0.03. The probability that the same gene would be found among the top 500 in both studies would thus be 0.03*0.03 = 0.0009 (about 0.1%), or 0.5 genes compared with the 77 found in our analysis.

At the pathway level, 10 of 18 pathways, or 56%, were the same in the NP-nMPC and nMPC-MPC transitions. Overlap at the functional level was even stronger: 18 of 21, or 86%, of functional categories significant in the transition from NP to primary nMPC were also significant for the transition from nMPC to MPC.

We looked at the correlation between gene expression and Gleason score by using data from four studies 41424344. We found that expression of ITGA5 and ITGAL were significantly negatively correlated with Gleason score, with corresponding P values of 0.003 and 0.02. We also considered the correlation of the expression of integrin ligands (that are largely components of the extracellular matrix) with the Gleason score. We found that the expression levels of seven integrin ligands (COL4A6, COL13A1, FGB, COL19A1, COL18A1, COL14A1, and COL11A2) were negatively correlated with the Gleason score, suggesting that the suppression of integrins and/or their ligands is associated with more advanced tumor grade.