Gene expression analysis reveals a pitfall in the molecular research of prostate tumors relevant to Gleason score

Abstract

Gleason score (GS) is a powerful prognostic factor in prostate cancer (PCa). A GS-7 tumor typically has the primary Gleason (architectural) pattern and secondary prevalent one being graded with 3 and 4 (or 4 and 3), respectively. Due to the well-known intratumoral multifocal occurrence of different patterns, a biological sample from a GS-7 tumor used in a molecular experiment will be uncertain regarding the actually represented pattern if no special attention is given to specimen preparation. In this study, by an integrative analysis of several published gene expression datasets, one of which is the profiling of the paired GP-3 (Gleason pattern 3) and GP-4 (Gleason pattern 4) specimens of 13 GS-7 tumors, we demonstrate that such an uncertainty can be frequently observed in the published data. More specifically, our results suggest that the GS-7 specimens used to generate the frequently-cited The Cancer Genome Atlas (TCGA) data and the Gene Expression Omnibus (GEO) dataset GSE21032 which largely are individual GP-3 or GP-4 specimens rather than the “intermediate” specimens of GP-3 and GP-4. This indicates a pitfall in the existing molecular research of prostate tumors relevant to GS and in GS-related molecular biomarker identification using the previously documented data.

1. Introduction

Prostate cancer (PCa) is the most prevalent non-skin cancer type in American men. Gleason score (GS) is a powerful prognostic factor for this cancer type. The determination of GS is based on the Gleason (architectural) patterns of a tumor specimen. The primary and secondary most prevalent patterns are first assigned a grade from 1 to 5 by pathologists. Then a GS is calculated by the sum of these two grades.

Adenocarcinomas amounts to 95% of PCa cases (https://www.cancer.gov/types/prostate/hp/prostate-treatment-pdq#cit/section_1.21). Over a half of them have the GSs equal to 7.^1–3 A GS-7 carcinoma typically has the primary Gleason pattern and secondary prevalent one being graded with 3 and 4 (or 4 and 3), respectively. While it is likely that a Gleason pattern 3 (GP-3) and a Gleason pattern 4 (GP-4) have a common precursor, they subsequently undergo substantial divergence in genetics, resulting in distinguished mutation spectra.^4,5 A GP-3 cannot directly progress into a GP-4, in general.^4,6

Due to the well-known intratumoral multifocal occurrence of different Gleason patterns,⁷ a biological sample from a GS-7 PCa tumor used in a molecular experiment may be uncertain regarding the actually-represented pattern unless special attention is given to specimen preparation. Unfortunately, this issue did not receive sufficient attention from most of the GS-relevant molecular research. Such ignorance could compromise the validity of the obtained results.

In this study, by an integrative analysis of several publicly-available gene expression datasets, we demonstrate that the uncertainty of architectural patterns in GS-7 PCa specimens is never unusual in the previous studies. This finding is of significance for deriving biological/clinical implications from the documented molecular data and for designing new experiments.

2. Material and Methods

2.1. Data

Three datasets are analyzed in this study. The first, i.e. GSE52560, is a unique dataset that contains the gene expression profiling of the paired GP-3 and GP-4 specimens of 13 GS-7 prostate carcinomas, measured with Affymetrix Human Exon 1.0 ST Arrays.⁴ We downloaded the raw data from GEO and generate three gene expression matrices (i.e. GSE52560-M-1, -2, -3) using the different normalization/summarization pipelines. GSE52560-M-1 is obtained by applying the frozen Robust Multi-array Analysis (fRMA) algorithm.⁸ The expression levels of 22011 Affymetrix core feature-sets (genes) are calculated by the rma() function in the R package “frma”, with the default arguments used in the implementation. GSE52560-M-2 and GSE52560-M-3 are generated by the application of Single-Channel Array Normalization (SCAN) and Universal exPression Codes (UPC) algorithms,⁹ respectively. The expression levels of 111655 RefSeq genes (including mRNAs, non-coding RNAs and computationally annotated RNAs) are calculated by the SCAN() or UPS() functions in the R package “SCAN.UPC”, with the default arguments. The second contains the log2 transformed level-3 RNA-seq gene expression profiling of 333 TCGA prostate adenocarcinomas (PRAD) samples whose Gleason patterns were reviewed by two pathologists.² A refined version of this dataset, containing 65 GS-6 (3 + 3), 102 GS-7(3 + 4), 78 GS-7(4 + 3) and 44 GS-8(4 + 4) tumors, is focused in our study. The third is a section of the dataset GSE21032.¹ It contains 130 samples with GS information, among which the numbers of GS-6(3 + 3), GS-7(3 + 4), GS-7 (4 + 3) and GS-8 and 9 (4 + 4 or 4 + 5) tumors are 41, 53, 21 and 15, respectively. The expression profiling was measured with Affymetrix Human Exon 1.0 ST Arrays and was pre-processed using Aroma Affymetrix software. We repeatedly apply the cohort-wide quantile normalization to each of these five gene expression matrices before further analyses.

2.2. Statistical analysis

The paired t-test is used to identify the genes which are differently expressed between the GP-3 and GP-4 specimens. The independence of two different sample partitions is tested by the Fisher’s method. The Ward method, with Manhattan distance, is used in the hierarchical clustering analysis via the R “stats” function hclust(). The clustering analysis results are visualized by the function heatmap.2() in the R package “gplots”.

To evaluate the classification strength of the expression profiling of a gene set for distinguishing the GS-8(4 + 4) tumors from the GS-6(3 + 3) ones, we use the support vector machine (SVM),¹⁰ leave-one-out cross validation, Fisher’s exact test and receiver operating characteristic (ROC). First, based on the Gleason patterns, the N tumors in a cohort are divided into two classes, e.g. 3 + 3 (i.e. “−1” group) and 4 + 4 (i.e. “1” group). The labels of these tumors are then saved in a vector Y = (y₁, y₂, … y_i, … y_N), where y_i ∈ (−1, 1). After that, the (assumedly unknown) class of a leave-out tumor i is predicted using its gene expression profiling $\left({\stackrel{\rightharpoonup }{x}}_i\right)$ by the SVM model, which is trained on the data $\left(\left\{{\stackrel{\rightharpoonup }{x}}_j;y_j\right\}\right)$ of the other N − 1 samples. That is, we have

$$\begin{array}{ccc}{\widehat{z}}_i=\text{sign}\left(t_i\right),& t_i=\sum _{j\in S}^{j\ne i}{a}_j{y}_{j}k\left({\stackrel{\rightharpoonup }{x}}_j,{\stackrel{\rightharpoonup }{x}}_i\right)+b,& S=\left\{1,2,3,\dots ,N\right\}\end{array}.$$

In the equations, ${\widehat{z}}_i$ denotes the predicted category (1 or −1) for the ith sample; t_i is the decision value, $k\left({\stackrel{\rightharpoonup }{x}}_j,{\stackrel{\rightharpoonup }{x}}_i\right)$ is the kernel function, and {a_j} and b are the model parameters decided in the previous training process. Third, by summarizing the true label vector Y and the predicted label vector $\check{Z}=\left({\widehat{z}}_1,{\widehat{z}}_2,{\widehat{z}}_3,\dots {\widehat{z}}_N\right)$, a 2 × 2 contingency table is generated, on which the Fisher’s test of independence is performed. Meanwhile, the sensitivity and specificity are calculated from the table. Finally, by combining the true label vector Y and an assemblage of tumor sample classifications based on the decision value vector T = (t₁, t₂, t₃, … t_N) and using serially changed cutoffs, a ROC is generated and the area under curve (AUC) is calculated. A SVM model is created by the svm() function in the R package “e1071”.

3. Results

Our analysis moves towards revealing molecular evidence for the hypothesis that, in typical previously-reported genome-wide research on PCa, the set of mini-dose experimental specimens blindly sampled from the prostatectomy specimens of GS-7 tumors is largely composed of either GP-3 specimens or GP-4 specimens. Here, the term “mini-dose” indicates that the body of a specimen is small and “blindly” indicates that the Gleason pattern of the mini-dose specimen is not specified in specimen preparation or is not appropriately recorded.

3.1. Architectural pattern-related genes (apr-Genes)

Using each of the three expression matrices (i.e. GSE52560-M-1, −2, or −3) derived from the GSE52560 data, we generate a list of genes that are up- or down-regulated in GP-4 specimens compared to GP-3 specimens by the criterion of p-value < 0:05. Of the three lists, 303 genes are consistently labeled as “down-regulated” or “upregulated”. In other words, each one of these genes is commonly determined as an upor down-regulated gene from the three expression matrices. Hereafter, we consider them as architectural pattern related genes (apr-genes). Based on GSE52560-M-1, all these genes have the log2 transformed expression levels over 1.0 in every sample, and 90% of them have the between-groups (GP patterns) log2 fold changes ≥ 0.29 or ≤ −0:29. Using these genes’ expression profiling in GSE52560-M-1, we partition the 26 specimens into two clusters, which correspond to the GP-3 and GP-4 groups (Fig. S1) with a 23% misclassification rate.

We perform functional enrichment analysis of the 303 significant genes.¹¹ Six gene ontology (GO) terms are over-represented by the down-regulated genes (N = 96) with the BH adjusted p-value less than 0.05. They include CC: GO:0031012~extracellular matrix, BP: 0006936~muscle contraction, MF:0008307~structural constituent of muscle, CC:0005925~focal adhesion, MF:0003779~actin binding and CC:0030018~Z disc. With the same cutoff of adjusted p-values, only one GO term, i.e. CC:0070062~extracellular exosome, is over-represented by the up-regulated genes (N = 207).

3.2. GS-6(3+3) versus GS-8(4+4) resembles to GP-3 versus GP-4

Of the 303 apr-Genes, 288 are included in the TCGAs level-3 RNA-seq data. The result from machine learning analysis shows that expression profiling of these genes is a strong prognostic signature for distinguishing the GS-6(3 + 3) samples from the GS-8(4 + 4) ones, with the AUC being up to 96% (Fig. 1(a)). Also, the result of biclustering analysis suggests that the co-expression patterns of the apr-Genes in the aggregate of GS-6(3 + 3) and GS-8(4 + 4) samples are similar to those presented in the aggregate of GP-3 and GP-4 specimens (Fig. 1(b)). This perception can be further scrutinized from two aspects. On one hand, if we divide the 109 tumors into five clusters, only 13% of GS-6(3 + 3) tumors will be assigned to the two clusters which contain 71% of GS-8(4 + 4) tumors. On the other hand, if we divide the 298 genes into two clusters, the first will nearly exclusively contains the genes that are down-regulated in GP-4 specimens (compared to GP-3 specimens) while the second is predominated by the genes that are up-regulated in GP-4 specimens. These observations imply that the differences of transcription program between the GS-8 (4 + 4) and GS-6(3 + 3) samples resemble to the differences between the GP-3 and GP-4 specimens.

Fig. 1.

Machine learning and clustering analysis of the expression quantities of GS-6(3 + 3) and GS-8 (4 + 4) PCa specimens in the TCGA data. (a) The expression profiling of 298 architectural (Gleason) pattern-relevant genes (apr-Genes) is a prognostic signature for distinguishing the GS-6(3 + 3) samples from the GS-8(4 + 4) ones. In the statistical analysis, GS-8(4 + 4) tumors are considered as “positive” cases. The printed Sn (sensitivity), Sp (specificity) and p-value are calculated from a contingency table determined by the signs (+ or −) of the decision values of tumor cases. The ROC curve is generated from a set of contingency tables which depend on the signs of the differences between the decision values of tumor cases and serially-changed cutoffs. (b) Co-expression patterns of apr-Genes. In the color bar below the top dendrogram, blue and orange indicate the GS-6(3 + 3) and GS-8(4 + 4) samples, respectively. In the color bar beside the left dendrogram, green and red, respectively denote the genes that are up-regulated or down-regulated in the GP-4 specimens compared to GP-3 specimens, as demonstrated by the GSE52560 data as well.

3.3. GS-7 resembles to the aggregate of GS-6(3+3) and GS-8(4+4)

Another implication of the aforementioned results is that the samples of GS-6(3 + 3) tumors or GS-8(4 + 4) tumors are similar to GP-3 specimens or GP-4 specimens in transcriptomic profiles, respectively. In other words, a set of samples collected from the GS-6(3 + 3) and GS-8(4 + 4) tumors resembles to a mixture of the GP-3 and GP-4 specimens. Based on this indication, we can state that if the co-expression patterns of apr-Genes in GS-7(3 + 4 or 4 + 3) samples are similar to the patterns present in the aggregate of GS-6(3 + 3) and GS-8(4 + 4) samples, it will provide supporting evidence for the aforementioned hypothesis about the constitution of a specimen set blindly sampled from the prostatectomy specimens of GS-7 tumors. The following analysis is guided by this reasoning.

We first refine the list of apr-Genes by excluding those that cannot be verified using the data of the TCGA GS-8(4 + 4) and GS-3(3 + 3) samples. Then, the biclustering analysis is individually performed on the expression matrices of the remaining 81 genes (Table S1) for three tumor sets (Set-1, Set-2, Set-3). Set-1 includes 65 GS-6(3 + 3) and 44 GS-8(4 + 4) tumors, Set-2 includes 102 GS-7(3 + 4) tumors, and Set-3 includes 78 GS-7(4 + 3) tumors. As shown in Figs. 2(a)–2(c), similar co-expression patterns are observed in the apr-Genes of these three sets. In particular, the samples in Set-2 (and Set-3) are partitioned into two clusters whose gene (co-) expression patterns resemble to the two clusters of Set-1, in which GS-6 (3 + 3) samples (corresponding to GP-3 specimens) and GS-8(4 + 4) samples (corresponding to GP-4 specimens) are predominant, respectively, with the p-value of the independence test less than 0.00001.

Fig. 2.

The resemblance of the co-expression patterns of 81 verified architectural (Gleason) pattern-relevant genes in three sample subsets of the TCGA prostate cancer (PCa) cohort. “3 + 3”, “4 + 4”, “3 + 4” and “4 + 3” indicate GS-6(3 + 3), GS-8(4 + 4), GS-7(3 + 4) and GS-7(4 + 3) tumor samples, respectively. The top dendrogram shows that tumors in each subset can clearly be partitioned into two clusters, which are marked by black and orange, respectively. In the color bar of the top dendrogram of the left heatmap, blue denotes a GS-6(3 + 3) tumor or a GP-3 pattern and orange denotes a GS-8(4 + 4) tumor or a GP-4 pattern. The consistent gray color bar of the top dendrograms in the middle and right heatmaps indicate the real-world practice that the Gleason patterns of the specimens actually used in the experiments are not recorded in the published TCGA data. In the color bars beside the left dendrograms of all the three heatmaps, green and red denote that the gene is up- and down-regulated in GP-4 specimens, respectively, in comparison to GP-3 specimens.

3.4. Repeatability

Using the same methods as described in Secs. 3.2 and 3.3, we analyze the GSE21032 data. Because the size (N = 7) of GS-8(4 + 4) group is too small to perform an effective ROC-AUC analysis, we combine the 8 GS-9(4 + 5) samples with the GS-8 group. This procedure is appropriate because both GP-5 and GP-4 are histological markers for high-grade PCa. As shown in Figs. S2 and S3, the similar patterns of resemblance identified in TCGA samples are also observed in this data. Specifically, first, the expression profiling of the apr-Genes shows a high predictive strength for distinguishing the GS-6(3 + 3) samples from the GS-8(4 + 4) and GS-9(4 + 5) ones (AUC = 0:95). Second, the stratification of co-expression patterns of these genes shown in the groups of GS-7(3 + 4) and GS-7(4 + 3) is similar to that observed in the aggregated category of GS-6(3 + 3), GS-8(4 + 4) and GS-9(4 + 5) samples. In addition, Fig. S3 implies that the down- and up-regulation profiles of apr-Genes determined from comparing GP-4 and GP-3 are largely verifiable in the comparison of GP-3 and GP-5 because most of the GS-9(4 + 5) tumors fall into the cluster of GP-8(4 + 4) tumors.

4. Discussion

In this study, we provide evidence for the hypothesis that the set of experimental specimens of GS-7 tumors actually used in typical genome-wide research during past years are largely composed of either GP-3 specimens or GP-4 specimens. Because the analyzed gene expression datasets include hundreds of samples and are published in top scientific journals, it is not too bold to conclude that the uncertainty of architectural patterns in GS-7 PCa samples is never uncommon in previous studies. This insight indicates a pitfall in the current molecular research of PCa relevant to GS.

As an intermediate step towards the main objective of this study, we demonstrate that the architectural pattern related genes (apr-Genes) constitute a strong prognostic signature for distinguishing the GS-6(3 + 3) samples from the GS-8(4 + 4) ones. This is somewhat different from the result of True et al.¹² Applying their pinpointed signature (i.e. 86 genes) to an independent tumor set, True et al. found that 11 of 12 GS-6(3 + 3) samples and 5 of 10 GS-8(4 + 4) samples could be correctly classified. In their opinion, such a high misclassification rate for GS-8(4 + 4) specimens suggests that GP-4 cancers are diverse and, in some cases, exhibit molecular features common to GP-4 cancers.

Contemporary analyses of the information of patients with organ confined, pure GS-6(3 + 3) tumors indicate favorable outcomes, with a negligible 10-year biochemical recurrence (BCR) rate and no prostate cancer-specific mortality (PCSM).^13,14 The risk of BCR and PCSM increases with the climbing proportions of the GP-4 and GP-5 component, from 3 + 3 with tertiary 4 to 3 + 4, 4 + 3 and 4 + 4, in the prostatectomy specimen.^6,15,16 While the uncertainty of the Gleason pattern in the specimen of a GS-7 tumor actually used for a molecular test does not impact the value of the GS system in predicting the behavior of PCa, it does impede the understanding of the underlying biology and the identification of molecular markers. For example, the GP-3 and GP-4 components of a GS-7 tumor may be driven by different sets of somatic mutations and epigenetic events. The molecular analysis of a GP-3 (or) GP-4 specimen from the tumor cannot reveal such complexities. Especially, when the measured specimen comes from a GP-3 structure of a GS-7 tumor, the risk factors for potential malignant growth, which more likely exist in GP-4 than in GP-3 tumors, could not be pinpointed. Such a structural uncertainty in specimens also limits the utility of available high-throughput data of GS-7 tumors in other areas of basic and translational research. For example, when one wants to compare the high-grade PCas between two populations, the data of a GS-7 specimen with uncertain Gleason patterns cannot be used with confidence. This could further intensify the problem already present from the shortage of data for minority populations, such as African Americans whose PCa occurrence and mortality rates are much higher compared to other populations.^17–19

An effective solution to the uncertainty regarding the actually represented pattern of a GS-7 tumor and the potential impacts on molecular clinical researches is first to “predict” the Gleason pattern of the mini-dose experimental specimen of a GS-7 tumor using machine learning methods, and then to compare the GP-3 and GP-4 specimens (including the predicted ones) to pinpoint architectural pattern-relevant genomic, epigenetic and transcriptomic features. Such a solution would function as a catalyst in solving the aforementioned data shortage problem faced by PCa health disparities research.

Biography

Wensheng Zhang is a bioinformatics scientist at Xavier University of Louisiana, USA. He earned his Ph.D. in statistical genetics and poultry science from the University of Georgia in 2004. His research interests include, but not limited to, the statistical methods for microarray and next-generation sequence data, the integration of genomic information of different kinds and sources, the integration of linear model methodology and machine learning algorithms on the analysis of genomic data, and the applications in cancer biology and clinical oncology. He has published over 30 first-authored papers in peer reviewed journals and proceedings.

Yan Dong is a Professor in the Department of Structural and Cellular Biology at Tulane University School of Medicine. She has a long-standing interest in prostate cancer particularly in altered androgen receptor signaling in castration-resistant prostate cancer and mechanisms and efficacies of candidate prostate cancer interventions. Despite the advent of new hormone therapy drugs for metastatic castration-resistant prostate cancer, disease progression after hormone therapies remains the most critical challenge in the clinical management of prostate cancer. Emerging evidences indicate that prostate cancers can adapt to all current hormone therapies by signaling through constitutively-active androgen receptor splice variants that lack the ligand-binding domain. We are currently characterizing how these splice variants are generated and their mechanisms of action.

Kun Zhang is a professor of Computer Science at Xavier University of Louisiana. Her research interests include bioinformatics, machine learning and big data analytics. Her collaborative work with others received the IEEE ICDM’06 Best Application Paper Award, IEEE ICDM’08 Contest Crown Award and IEEE ICDM’10 Top-10 Data Mining Case Studies. She has published over 70 peer reviewed articles in Nucleic Acids Research, Bioinformatics, and among others. Her lab is working on developing machine learning and multi-omics data integration algorithms to facilitate translational precision oncology research.