Validation of proposed prostate cancer biomarkers with gene expression data: a long road to travel

The analysis of the literature has led to the identification of over 20,000 articles on prostate markers published between January 1, 2001 and June 1, 2011. Articles published in journals not indexed in the Journal Citation Reports and articles reporting on markers that are not measured as mRNA or protein expression were omitted from further analyses. Two hundred forty-four articles report for the first time at least one new mRNA or protein marker for a total of 380 markers. There is a trend towards slightly increasing numbers of articles reporting prostate cancer markers over time (Fig. 1a). The studies have been published in journals with a wide range of impact factors from 0.822 (Ca. J. Urol) to 18.97 (J. Clin. Oncol.) (Fig. 1b). The complete bibliography containing the list of references for all studies included in this analysis is available as Supplementary Table 1.

We analyzed all markers together irrespective of the potential application (diagnostic, prognostic, or follow-up) of the marker claimed in the original publication. We evaluated the markers for their diagnostic potential using a set of microarray data of the GeneSapiens database containing 147 normal prostate and 329 prostate cancer tissues. Of the 287 markers identified in the microarray data, 143 markers had significantly different expression values when normal and cancer tissues were compared (p < 0.01). Figure 2 shows the ten markers with highest (downregulated in cancer) and ten with the lowest (overexpressed in cancer) expression ratio and, for comparison, kallikrein-related peptidase 3 (KLK3), the gene encoding PSA. The lowest score, 0.46, is attained by filamin C (FLNC) [42] and the highest score, 3.00, by claudin 3 (CLDN3) [43] (KLK3/PSA = 0.56). The data for all markers are reported in Supplementary Table 2.

We investigated the expression of the markers in an independent dataset (GSE6919) containing data from 152 human prostate tissues including normal prostate tissue from healthy donors, prostate cancers, peritumoral tissues, and prostate metastases. Hierarchical clustering of the expression data for the 380 markers reveals that almost all of the metastases and many of the tumor tissues cluster together in a cluster distinct from the clusters containing mainly peritumoral and normal tissues, indicating that the combination of markers distinguishes to some extent healthy and tumor tissues. However, the clusters formed are not strongly distinct as the distances in the dendrogram are short (Fig. 3a). When the same analysis is limited to the 20 markers from Fig. 2 whose expression is most different in normal versus tumor tissues, the clusters formed become slightly more robust and all the metastases and the majority of tumors are in one cluster, yet the clusters are still not very distinct (Fig. 3b).

Figure 4 reports the expression scatter plots of the two markers with the strongest overexpression in cancer (alpha-methylacyl-CoA racemase/AMACR [44] and CLDN3 [43]) and in normal (FLNC [42] and keratin 5/KRT5 [45]) tissue and KLK3/PSA for comparison. The new markers do not appear to be clearly superior to KLK3/PSA inasmuch as their expression is not drastically different in normal and cancerous tissues and their expression in normal and cancer tissues varies widely not allowing for the classification of single patients according to the expression levels, although the expression differences are statistically significant.

The different significance level of the “new” markers as compared to KLK3/PSA (p = 1.2 × 10⁻¹⁶ for FLNC, 1.67 × 10⁻⁹ for CLDN3, 1.89 × 10⁻³ for KLK3/PSA) could indicate that these markers are more powerful for the discrimination of cancer and normal tissue. The potential as a diagnostic marker depends, however, at least for serum markers, also on the prostate-specific expression as compared to other tissues. We therefore used the GeneSapiens set of microarray data to monitor tissue-specific expression of the new markers as compared to KLK3/PSA. Figure 5 reports the expression patterns for the four best new markers and KLK3/PSA. This analysis shows that despite the greater expression difference in normal versus cancer tissues, the new markers are unlikely to be superior to KLK3/PSA given their widespread expression in other normal and neoplastic tissues as well as in tissues affected by other diseases. The prostate specificity of KLK3/PSA is unmet.

We next asked whether the markers identified have any prognostic potential. We used the GSE16560 dataset of 281 prostate cancers. The samples are derived from FFPE material from transurethral resection of prostate at the time of the initial diagnosis. Patients who died of the disease within 10 years (n = 140) and patients who survived at least 10 years (n = 141) were selected for the analysis allowing for a clear-cut distinction. For 280 of the 380 proposed markers, a corresponding probeset could be identified on the arrays used for this study. Hierarchical clustering of the gene expression data of these markers did not show strong associations of gene expression values with status (lethal or indolent) or Gleason score (Fig. 6).

In order to identify the prognostic potential of single markers, we performed Cox regression analyses using the same dataset. Figure 7 shows Kaplan–Meier survival curves for two markers, BIRC5/survivin [46] and NKX3-1 [47, 48], among those with the lowest logrank test p value (p = 0). Low-risk and high-risk cases show a significantly different survival, yet it is unlikely that differences as observed here could guide treatment decisions or follow-up screenings. The collection of Kaplan–Meier curves for all markers analyzed is available as Supplementary Fig. 1.

Multigene signatures have been shown to have a considerable prognostic power for several cancers [49]. We therefore asked whether a multivariate score of the markers that are significantly differentially expressed between low- and high-risk cases has a clinically relevant prognostic power. The multivariate model was calculated in a backward manner in order to leave as many genes in the model as possible. The genes selected for the model are the ATP-binding cassette (ABC) transporter with unknown substrate and function (ABCA5) [50], the engrailed homeobox gene 2 (EN2) [51], the 17-beta-hydroxysteroid dehydrogenase type 3 that converts androstenedione to testosterone (HSD17B3) [52], the NK3 homeobox 1, a negative regulator of epithelial cell growth in prostate tissue, (NKX3-1) [47], the signal transducer and activator of transcription 6 that mediates the anti-apoptotic effects of interleukin 4 (STAT6) [53], the E2F transcription factor 1 that is involved in the control of cell cycle progression (E2F1) [54], the folate hydrolase (prostate-specific membrane antigen) 1, a glutamate carboxypeptidase (FOLH1) [47], the proteasome (prosome, macropain) subunit, alpha type, 7 that plays a role in the cellular stress response by regulating hypoxia-inducible factor 1 alpha (PSMA7) [55], and the topoisomerase (DNA) II alpha (TOP2A) [54]. Table 1 shows the results of the multivariate analysis. A score indicating the strength of correlation between the expression of the given gene and survival (column B) is calculated for each gene. The global multigene score (MGS) is obtained by the sum of the expression values multiplied by the score assigned. The median value of the score is then used to classify the samples in low and high risk. Kaplan–Meier survival curves for the commonly used Gleason scoring system (Fig. 8a), the presence or absence of the fusions involving the V-Ets erythroblastosis virus E26 oncogene homolog (ERG; Fig. 8b), and the multigene score are plotted (Fig. 8c). All three prognostic measures yield risk classes with significantly different risk of death from prostate cancer (logrank test p = 0). The discrimination of high- and low-risk groups using the multigene score (Fig. 8c) is clearly superior to that observed for single genes (see Fig. 7 and Supplementary Fig. 1). The survival differences of the low- and high-risk groups are evident from the very beginning of follow-up. After 5 years, approximately 86 % of the low-risk cases and 56 % of the high-risk cases are alive, and after 10 years, these figures become 73 and 24 %, respectively. This analysis shows that the combined score can distinguish prostate cancer patients with a significantly different risk of death of prostate cancer, similar to what is obtained by Gleason scoring (Fig. 8a). ERG fusions found in 46 cases (226 cases without fusion, for 9 cases the fusion status is unknown) also confer a bad prognosis, yet the absence of a fusion is not a good indicator of an indolent evolution of the cancer (Fig. 8b).

The Gleason scoring system [56] is commonly used for prostate cancer prognosis. The combination of new molecular markers with the Gleason score must be assessed. We therefore calculated Kaplan–Meier survival curves for cases with a Gleason score below 7 or equal to 7 (=3 + 4) and cases with Gleason > 7 or equal to 7 (=4 + 3). The application of the combined marker score to the former cases shows an improved distinction of low- and high-risk cases (Fig. 8d) and cases with low Gleason score and low-risk multigene score show a clear difference in survival from the very beginning of follow-up as compared to cases where both scores indicate high risk. Cases with low and high Gleason score are further divided by the application of the MGS. This creates two intermediate groups (Gleason low, MGS high; Gleason high, MGS low) with similar Kaplan–Meier curves (Fig. 8d). Combination of the MGS with ERG fusion status shows that MGS low-risk cases without fusion have a good prognosis and cases with MGS high risk or the ERG fusion, or both, have a bad prognosis (Fig. 8e). Cases can be classified considering at high risk those cases that receive an indication of high risk by either the Gleason score or the MGS thus considering the intermediate cases as high risk. In this way, an additional 40 of 162 cases with Gleason 6 or 7 (=3 + 4) would be correctly identified as high-risk cases at the expense of 20 indolent cases that would be considered at high risk (Table 2). If instead Gleason 7 would be considered high risk with no regard of the status of the major and minor components of the tumor, additional 39 would be correctly identified as high risk yet 40 actually indolent cases would be classified at high risk. The combination of the MGS with Gleason can therefore be expected to slightly improve the assessment of cases with Gleason score of 6 or 7 (see also Table 2).

Calculation of multigene scores often leads to over-fitting yielding scores that strongly depend on the specific dataset on which they have been calculated. To avoid over-fitting, the dataset must be randomly divided into a training set on which the score is calculated and a test set to which the score is applied [57]. When we applied this strategy to the selected prostate biomarkers, the resulting risk classes show a significantly different risk (logrank test p = 0) in the training set. The same score yields a similar distinction when applied to the test set that also is statistically significant (logrank test p = 0.005) (Supplementary Fig. 2).

We further validated the MGS on an external dataset (GSE21034). The application of the MGS to this dataset also yielded risk classes with significantly different risks (p = 0.002; Fig. 8f). This dataset is based on a completely different array type (exon arrays) and probe design. Thus, the application of the score calculated on the data derived from a different platform can lead to an underestimation of the discrimination power of the classifier.

Finally, we asked whether the many biomarkers identified are functionally interrelated or independent. For this purpose, we performed a correlation analysis using more than 10,000 microarray gene expression data sets. Figure 9 shows the correlation heat map. There is generally a considerable correlation (r > 0.5) of any marker with several others. In order to understand whether correlated markers belong to groups of genes that exert similar functions or participate in similar biological processes, we analyzed the enrichment of gene ontology terms using the Database for Annotation, Visualization and Integrated Discovery [58] for the four predominant, yet arbitrarily selected clusters of the correlation map. Cluster 1 (Fig. 9) shows enrichment of several angiogenesis-related gene ontology (GO) annotations; most of which contain the angiogenic factors VEGFA and VEGFC as well as HGF. Cluster 2 shows an enrichment of extracellular matrix-related GO terms dominated by several matrix metalloproteinases (MMP2, MMP9, and MMP13). Cluster 3 shows GO terms related to peptidase activity containing several kallikreins and in cluster 4 epithelial–mesenchymal transition-related GO categories predominate (see Supplementary Table 3 for complete data). These four biological processes are clearly related to cancer development and progression. Interestingly, cell growth and proliferation are not among the most enriched GO terms despite the important role of cell proliferation in cancer prognosis.