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

Abstract

Biomarkers are important for early detection of cancer, prognosis, response prediction, and detection of residual or relapsing disease. Special attention has been given to diagnostic markers for prostate cancer since it is thought that early detection and surgery might reduce prostate cancer-specific mortality. The use of prostate-specific antigen, PSA (KLK3), has been debated on the base of cohort studies that show that its use in preventive screenings only marginally influences mortality from prostate cancer. Many groups have identified alternative or additional markers, among which PCA3, in order to detect early prostate cancer through screening, to distinguish potentially lethal from indolent prostate cancers, and to guide the treatment decision. The large number of markers proposed has led us to the present study in which we analyze these indicators for their diagnostic and prognostic potential using publicly available genomic data. We identified 380 markers from literature analysis on 20,000 articles on prostate cancer markers. The most interesting ones appeared to be claudin 3 (CLDN3) and alpha-methysacyl-CoA racemase highly expressed in prostate cancer and filamin C (FLNC) and keratin 5 with highest expression in normal prostate tissue. None of the markers proposed can compete with PSA for tissue specificity. The indicators proposed generally show a great variability of expression in normal and tumor tissue or are expressed at similar levels in other tissues. Those proposed as prognostic markers distinguish cases with marginally different risk of progression and appear to have a clinically limited use. We used data sets sampling 152 prostate tissues, data sets with 281 prostate cancers analyzed by microarray analysis and a study of integrated genomics on 218 cases to develop a multigene score. A multivariate model that combines several indicators increases the discrimination power but does not add impressively to the information obtained from Gleason scoring. This analysis of 10 years of marker research suggests that diagnostic and prognostic testing is more difficult in prostate cancer than in other neoplasms and that we must continue to search for better candidates.

Fig. 1

Publications on prostate cancer biomarkers 2001–2011. a Publications per year. b Distribution of publications according to impact factor

Fig. 2

Expression of prostate cancer biomarkers in healthy and tumoral prostate tissues. The ratios of expression in healthy and tumoral prostate tissues of the 20 prostate cancer biomarkers that are most significantly differentially expressed are reported. KLK3 (PSA) has been added for comparison

Fig. 3

Hierarchical clustering marker gene expression in human prostate tissues from dataset GSE6919 using Euclidean distance measures and average linkage. The state of the tissue is indicated by a color code in the bar above the dendrogram (green = prostate tissues from healthy donors, yellow = peritumoral tissue, orange = tumor tissue, red = metastases). For markers represented by more than one probe set on the array, all probe sets were included in the analysis. a All prostate cancer biomarkers. b The 20 best markers from Fig. 2

Fig. 4

Expression scatter plots of the four best prostate cancer biomarkers (claudin 3 (CLDN3), alpha-methylacyl-CoA racemase (AMACR), keratin 5 (KRT5), filamin C (FLNC)) in comparison to KLK3 (PSA). Expression data for healthy, peritumoral, tumoral, and metastatic prostate tissues are shown for the four most differentially expressed markers in comparison to KLK3 (PSA)

Fig. 5

Relative expression of prostate cancer biomarkers in various tissues a claudin 3 (CLDN3), b alpha-methylacyl-CoA racemase (AMACR), c filamin C (FLNC), d keratin 5 (KRT5), e KLK3/PSA. Note that only KLK3/PSA is highly specific for prostate tissues

Fig. 6

Hierarchical clustering of 281 prostate cancer tissues. Gene expression values of the genes encoding potential prostate cancer biomarkers in 281 prostate cancers from dataset GSE16560 were clustered using Euclidean distance measure and average linkage. Cancer status (indolent = white, lethal = black) and Gleason score (5 = green, 6 = yellow, 7 = orange, 8 = pink, 9 = red) are indicated in the bars above the dendrogram

Fig. 7

Kaplan–Meier survival analysis of prostate cancer biomarkers. Kaplan–Meier curves for the two markers with the most significant prognostic potential based on data from dataset GSE16560 are shown

Fig. 8

Multivariate models for prostate cancer prognosis. Prognostic prostate cancer biomarkers were combined in a prognostic multigene model using multivariate Cox regression analysis and dataset GSE16560 (see also Table 1). a Kaplan–Meier survival analysis applying Gleason score (low risk = <7 or 7 (=3 + 4), high risk = >7 or 7(=4 + 3)). b Kaplan–Meier survival analysis for cases with and without rearrangements of the gene ERG. c Kaplan–Meier survival analysis for the multigene score; cases are assigned according to the median of the score. d Combination of Gleason score with the multigene score (assignment of cases as above). e Combination of ERG fusion status and multigene score. f Application of the model to the external dataset GSE21034. The scores calculated on GSE16560 were directly applied

Fig. 9

Correlation map of prostate cancer biomarkers. The expression correlation of the prostate cancer biomarkers is calculated and plotted as a heat map. Strong correlations are indicated by a color code (blue, <−0.5; red, >0.5). Arbitrarily selected clusters containing markers with high correlation are indicated by black squares (1–4): Cluster 1 – enrichment of angiogenesis-related genes; Cluster 2 – enrichment of extracellular matrix- and matrix metalloproteinases; Cluster 3 – enhanced peptidase activity and kallikreins; Cluster 4 – enrichment for epithelial–mesenchymal transition