Chromosomal instability in untreated primary prostate cancer as an indicator of metastatic potential

We received institutional review board (IRB) approval to abstract data and procure both fresh and archival PNBX samples from men diagnosed and treated within the Greater Los Angeles VA (GLA-VA) Healthcare System (protocol numbers PCC2018–020201 and PCC2010–11489). Informed consent was obtained for all prospectively collected tissue specimens. A consent waiver was approved for collection of archival samples and data. All specimens were stripped of personal health information and identifiers.

After Institutional Review Board approval was obtained, 99 PC cases with archival biopsy tissue available were selected for RNA sequencing (RNAseq) from a total cohort 1927 cases identified within the GLA-VA cancer registry or procedure logs. The selected cases were divided into sub-cohorts based upon tumor burden at diagnosis or recurrence/progression documented in imaging reports, including ^99mTc-methylene disphosphonate (^99mTc-MDP) planar bone scintigraphy, ¹⁸F-NaF positron emission tomography (¹⁸F-NaF PET), ultrasounds, magnetic resonance imaging (MRI), computed tomography (CT) scans, and plain radiographs. Cases were designated clinical stage M1 if metastatic lesions were identified on imaging scans performed within 1 year of the diagnostic PNBX. Osteoblastic, osteolytic, and/or sclerotic bone lesions observed on bone scan were confirmed or by the presence of overlapping lesions on plain radiographs, CT or MRI imaging. Oligometastatic “oligo” disease was defined by < 5 extrapelvic lymph node and/or bone metastases and no visceral metastases. Polymetastatic “poly” disease was defined as > 5 metastases or any visceral involvement. M1 cases were designated M1-oligo or M0-poly based on the aforementioned tumor burden assessment. Cases that were considered M0 at diagnosis, but demonstrated eventual metastatic progression (M0-M) were designated M0-oligo or M0-poly based on tumor burden on imaging scans at the time of follow-up. Cases were designated as clinical stage M0 non-metastatic (M0-NM) if no metastatic lesions were identified at the time of last follow-up. If imaging scans were not performed, or equivocal results were obtained, the cases were categorized as MX and excluded from analysis. A significant subset of M0 cases from the original 1927 cohort did not have an indication for diagnostic imaging due to diagnosis of low- or intermediate-risk PC according to D’Amico classification [25].

Diagnostic hematoxylin and eosin (H&E) PNBX slides were reviewed by a genitourinary pathologist, and high-grade tumor areas (Gleason grade 4, 5, or neuroendocrine/small cell) were encircled. Transfer of annotated areas to the paraffin-embedded tissue blocks was performed and 1-2 mm sterile circular biopsy punches enabled manual procurement of formalin-fixed tissue from the block. A minimum of 2 (1 mm) cancer cores were used from each high-grade tumor area for RNA and/or DNA preparation. Total RNA was extracted from tissue cores using the Ambion Recover All Total Nucleic Acid Isolation Kit for FFPE (ThermoFisher). Samples were eluted with H₂0 and quantitated with nanodrop and bioanalyzer. Approximately 40-100 ng of RNA was used to generate the libraries using the TruSeq RNA Access Library Prep Kit (Illumina) according to manufacturer’s instructions. Quality control for libraries was performed with the HS DNA Qubit and bioanalyzer, followed by sequencing on the Hiseq3000 at 1 × 50.

DNA was extracted from 1 mm FFPE cores with AnaPrep 12 Automated Nucleic Acid Preparation Instrument (Biochain) and quantitated with Qubit (Life technology). Approximately 80 ng of gDNA was used in the Affymetrix OncoScan FFPE Assay kit (Affymetrix) according to manufacturer’s instructions. The Affymetrix GeneChip Scanner 3000 G7 was used and data was processed with Affymetrix Chromosome Analysis Suite 3.1 (ChAS 3.1) to generate probe level and gene level log2-ratio using normalized data.

Gene expression data and OncoScan CNA data were deposited into the GEO database with accession GSE147493 and GSE147353, respectively.

We computed scores for focal CNAs, broad CNAs, overall CNAs (broad and focal), TMB, genomic fusion, MSI, PC-CIN, and CIN70 using data from 473 prostate samples in TCGA. We used GISTIC 2.0 [26] to compute focal, broad and overall CNAs. CNAs with the GISTIC value less than − 0.3 were categorized as ‘loss’ and larger than 0.3 as ‘gain’. If the length of altered (loss or gain) regions was longer than 98% of the chromosomal arm, it was classified as broad CNAs, while less than 98% chromosomal arm were classified as focal CNAs. For each case, the broad CNA score was generated by calculating the total number of broad CNAs over the total number of chromosomal arms, and the focal CNA score was calculated based upon the number of focal CNAs over the total number of genes. MSI score was based upon the MANTIS scores as described in previously published reports [27]. We obtained masked copy number segment, Mutect2 result of somatic mutation from UCSC Xena Browser (http://​xena.​ucsc.​edu/​). Fusion data was obtained from a published report [28]. TMB score was calculated from counts of non-synonymous mutations in a sample divided by exon size per million bases (total exon size = 299.029409 MB). Fusion score was defined as the ratio of the number of fusion events and the total number of fusion types [28]. The signature activation scores were computed using a Z-score method [29].

For RNAseq data analysis, the quality of sequence reads from the RNAseq data were assessed and low quality reads were filtered using the FastQC tool (Babraham Bioinformatics, Cambridge, UK) and ShortRead (v. 1.30.0) package from R bioconductor (version 3.3). Quantification of gene level expression from preprocessed RNA-seq results were performed with the UCSC hg19 build of the Homo sapiens genome, through the use of the Subread aligner and the featureCounts software [30, 31]. To reduce systemic bias between samples, the Trimmed Mean Method (TMM) was applied to gene level expression counts [32]. Genes were filtered out and excluded from downstream analysis if they failed to achieve raw read counts of at least 2 across all the libraries. Differentially expressed genes (DEGs) were determined with false discovery rate (FDR) < 0.05 and fold change> = 1.5 obtained from the integrative hypothesis testing method [33]. In order to determine whether a set of genes showed statistically significant and/or concordant differences between two biological states such as M1 versus M0-NM, GSEA software tool was applied to RNAseq dataset. Briefly, gene sets were obtained from MSigDB [34] or previously published analysis [16] Genes in the RNAseq dataset were sorted in descending order using the ‘Signal2Noise’ ranking metric and computed enrichment score (ES) using a Kolmogorov-Smirnov running sum statistic for the gene set. Significance was computed using a null distribution of the ES generated from a random gene set by 1000 permutations. For the training and evaluation of the classifier using the gene signatures, two principal dimensions using principal component analysis (PCA) were extracted from expression matrix of the gene signature and then SVM algorithm was applied to determine the discrimination border between the two groups (M1-poly versus M0-NM).

The PNBX cohort (n = 99) was separated into five sub-cohorts based upon tumor burden at diagnosis (oligo or poly) or no tumor at diagnosis +/− metastatic progression (oligo, poly, or no metastases) on at last follow-up (Supplementary Table 1). Bivariable analyses compared patient clinical characteristics across the sub-cohorts using Chi-square (or Fisher’s exact) tests for categorical variables and ANOVA (or Wilcoxon rank-sum or median) tests for continuous variables. Kaplan-Meir survival analyses were performed for prostate cancer-specific and all-cause mortality stratified by tumor burden categories described above. In addition, BCR-free survival was examined in high and low risk groups defined by the risk index using Cox proportional hazard regressions of the gene signature. Clinical characteristics were also compared between CIN status (high v. low) stratified by patient race (African-American and white). Lastly, multivariate logistic regression analyses were conducted for CRPC, prostate cancer-specific mortality and all-cause mortality to calculate odds ratios for CIN status (high v. low), controlling for age, race, PSA (> 20 v. ≤20) and Gleason sum (> 8 v. ≤8). Analyses were performed using Stata (StataCorp LP, College Station, TX, USA), SAS 9.4 (SAS Institute Inc., Cary, NC, USA), MATLAB (Mathworks, Natick, MA, USA) and R (v.3.1 http://​www.​r-project.​org/​).

The CIN70 signature was previously derived by identifying differentially expressed genes in tumors displaying high versus low levels of chromosomal imbalance. CIN70 was applied to five CRPC transcriptome datasets [22, 35‐38]. The signature activation scores derived from primary tumors were compared to CRPC metastases (Fig. 1). CIN70 scores were significantly higher in CRPC metastases compared to primary tumors across all datasets.

Since the high CIN70 signature activation score strongly associated with CRPC metastases, we sought to determine if it correlated with genomic evidence of CIN (i.e., broad CNAs) as well as unfavorable tumor features and outcome in untreated PC tissue samples. To assess the correlations of specific types of genomic aberrations and CIN70 activation scores measured in primary PC, we employed the PC cohort (n = 473) found in TCGA (Fig. 2a). We considered five different types of genomic alterations including focal CNAs, broad CNAs, tumor mutational burden (TMB), gene fusions, and microsatellite instability (MSI) for each case. All genomic alterations were quantified using the individual scoring method (see Methods). Briefly, CNA events were identified in each sample and assigned broad or focal status based exclusively upon length using GISTIC 2.0 [26]. The TMB and fusion score were computed using the number of events identified in each case. The MSI score was based upon the MANTIS scoring method, and pre-computed scores for the PC cohort in TCGA were obtained from a previously published report [27]. Cases were displayed in a heat map based upon CIN70 score (high to low) in order to visualize recurrent genomic alterations associated with the CIN (Fig. 2a). Pathology Gleason score and biochemical recurrence (BCR) status were also included on the map for each case. Comparing CIN scores to specific classes of genomic alterations allowed determination of correlation coefficients (Fig. 2b). The CIN70 score most strongly correlated with broad CNAs (r = 0.52, Fig. 2c). There was a positive correlation with focal CNAs (r = 0.46), which was weaker with mutational burden (r = 0.34) and fusions (r = 0.31). No significant correlation between CIN70 score and MSI score was observed (r = 0.09, Fig. 2c). High CIN70 scores were associated with PC displaying Gleason score of 8 or higher (Additional File 1, Supplemental Fig. 1). PC cases displaying BCR were also more likely to display high CIN70 scores (Additional File 1, Supplemental Fig. 1). These findings confirm that high CIN70 score is reflective of broad CNA frequency in PC, which, in turn, is associated with aggressive disease and poor outcome in TCGA cases. To assess whether specific, recurrent CNAs differ between cases with high versus low CIN scores, TCGA cases classified as CIN70-high versus CIN70-low were compared (Fig. 2d). An increase in recurrent CNAs were identified throughout the genome, but no specific chromosomal locations where affected by increases in CNAs, confirming the global genomic impact of CIN.

Having established that CIN70 scores are highest in mCRPC and high-risk primary CSPCs contained in TCGA, we sought to evaluate transcriptomic profiles derived from untreated primary tumors of men diagnosed with de novo metastases (clinical stage M1). High-grade PC areas in PNBX were procured for RNA sequencing (RNAseq) from formalin-fixed and paraffin-embedded diagnostic PNBX of 99 patients (Fig. 3a). We accessed selected archival diagnostic PNBX from a racially and ethnically diverse cohort of 1927 men who were diagnosed and treated exclusively within a single Veterans Affairs (VA) healthcare system (Additional File 1, Supplemental Fig. 2). Sub-stratification of this cohort was performed based upon metastatic tumor burden (oligo versus poly) at diagnosis and follow-up (Additional File 1, Supplemental Fig. 2A). Kaplan-Meier curves demonstrated overall survival (Additional File 1, Supplemental Fig. 2B), with M1-poly and M0-poly cases displaying significantly shorter survival than M1-oligo and M0-oligo cases. For comparison, we selected high-grade PC cases without evidence of metastatic progression (M0-NM) over a median follow-up of 56 months. Clinical characteristics of the sequenced PNBX cases are provided in Additional File 1, Supplemental Table 1.

We aimed to identify the transcriptomic footprint of metastatic disease in primary tumors by comparing mCRPC and primary tumors collected from PNBX of men diagnosed with de novo metastatic disease (M1-oligo and M1-poly). We also questioned the amount of CIN in these cancers. Towards this goal, we identified 1234 DEGs by comparing RNA sequencing datasets of PCs from men in the VA cohort diagnosed with de novo metastatic disease (M1-poly or M1-oligo) versus non-metastatic (M0-NM) cases. We also derived DEGs through comparison of gene expression between primary tumors from RPs and metastases collected at rapid autopsy in two published mCRPC cohorts (Taylor and Grasso data sets) [22, 35]. Strong correlations were revealed between DEGs from the VA cohort and the DEGs derived from the Taylor (r = 0.64) or Grasso (r = 0.45) cohorts (Fig. 3b). Slightly weaker correlations were evident for M1-oligo cases (r = 0.34 and 0.27, respectively).

Next, the overlap amongst DEGs across datasets (PNBX, Taylor, and Grasso) was used to identify 157 shared DEGs (Fig. 3c). Functional enrichment analysis in the 157 DEGs demonstrated the greatest activity of pathways associated with mitotic nuclear division, cell proliferation and cell-cell signaling (Fig. 3d). A large portion of DEGs (89/157, 57%) had the same directionality in gene expression between primary tumors associated with de novo metastases and CRPC sampled at metastatic sites. In a multidimensional scaling diagrams supervised by the 157 DEG set, cases without metastases (M0-NM) and M1 cases (both M1-oligo and M1-poly) form distinct clusters (Fig. 3e). Similarly, when applied to the Taylor and Grasso cohorts, primary tumors were separated from metastases (Fig. 3e). Collectively, these results demonstrate that untreated primary tumors from men with de novo metastases possess gene expression profiles that are correlative to heavily treated mCRPC. These results suggest mCRPC biology is embedded in the primary tumors of M1 patients and has the potential to reveal biological mediators of metastases, castration-resistance and lethal PC at the time of diagnosis via standard-of-care PNBX.

We hypothesized that CIN70, which is highly elevated in mCRPC, may demonstrate a similar expression profile in M1 PNBX. Consequently, we evaluated enrichment of CIN70 gene expression via gene set enrichment analysis (GSEA) of M1 versus M0-NM cases. This analysis revealed significant up-regulation of CIN70 genes in M1 tumors (Additional File 1, Supplemental Fig. 3) [16]. In order for tumor cells to tolerate CIN, inactivation of the TP53 gene or its associated pathway is often required [39]. Interrogation of genes linked to the hallmark p53 pathway activation signature demonstrated down-regulated in M1 relative to M0-NM tumors (p = 0.031). Interestingly, there was no enrichment in the DDR gene signature, indicating that this potential mechanism of genomic instability may not be prevalent in de novo metastatic CSPC (Additional File 1, Supplemental Fig. 3).

Since M1-oligo and M1-poly cases displayed significant differences in PC-specific survival (Additional File 1, Supplemental Fig. 1B), we hypothesized that distinct biological triggers may influence tumor burden. Consistent with this idea, out of total 1234 DEGs in M1 versus M0-NM, a relatively small portion of the DEGs (105 genes; 9%) were common in both M1-oligo and M1-poly, while most of the genes (696/801 DEGs in M1-oligo and 433/538 DEG in M1-poly) were exclusively regulated in M1-oligo or M1-poly cases (Additional File 1, Supplemental Fig. 4A). These DEGs were grouped into 3 clusters (shared, oligo-dominant, and poly-dominant) based on their differential expression patterns (Additional File 1, Supplemental Fig. 4B). To identify cellular processes within each group, functional enrichment analysis was performed using DAVID software (Additional File 1, Supplemental Fig. 4C) [40]. While the oligo-dominant cluster was enriched in inflammatory response, steroid metabolic processing, cell-cell signaling, and cell differentiation, the poly-dominant cluster displayed the strongest enrichment in cell proliferation and mitotic cell division. Consistent with this, GSEA analysis of M1-poly versus M0-NM revealed a leading-edge subset of genes that significantly contributed to the enrichment of CIN70 and demonstrated significant up-regulation in M1 tumors (Fig. 4a and Additional File 1, Supplemental Fig. 3A). Notably, seven out of the top 19 leading-edge genes (PBK, CEP55, UBE2C, MELK, TPX2, PTTG1, and CDCA3) regulate mechanisms during mitosis [41]. We will refer to these seven genes as PC-CIN (prostate cancer-CIN). In order to determine whether the CIN70 signature genes or simplified PC-CIN associated with metastasis (M)-stage (M1 versus M0-NM) at diagnosis, we developed a prediction model using a support vector machine (SVM) algorithm (see Methods and Additional File 1, Supplemental Fig. 5) and tested its accuracy using PC-CIN or CIN70 genes. The model displayed a high level of accuracy in predicting metastasis stage with area under the curve (AUC) value of 0.90 for PC-CIN and 0.96 for CIN70 (Fig. 4b). Both CIN70 and PC-CIN appeared significantly enriched in mCRPC relative to primary tumors across five datasets (Fig. 4c and Additional File 1, Supplemental Fig. 5A, all p < 0.00001). PC-CIN activation score was also highest in Gleason 8 and higher PC samples in TCGA, as well as in cases of BCR (Fig. 4d-f).

To better understand the distribution of CIN70 and PC-CIN scores in the context of the PNBX cohort (n = 99), we created an integrative heat map of CIN70 genes split into functional groups, as well as the 7 PC-CIN genes (Fig. 5a). Embedded in this heat map is the CIN70 score, with cases arranged from CIN70-low to CIN70-high, the disease stage (M0-NM, M0-oligo, M0-poly, M1-oligo, M1-poly), and the Gleason Sum (6–10). The heat map allows observation of the pattern of distribution of de novo metastatic (M1) cases along the spectrum of CIN scores. Interestingly, a bimodal distribution of M1 cases is observed, with 23/63 displaying CIN70 scores in the lowest third and 25/63 displaying CIN70 scores in the highest third. PC-CIN gene expression variability appeared to mirror the expression pattern of CIN70 genes. A volcano plot of PC-CIN genes in M0-NM versus M1-poly cases in the PNBX cohort demonstrates differential expression (Fig. 5b). PC-CIN scores are significantly higher in M1-poly cases compared to M0-NM cases (p = 0.0426), however, the wide range of PC-CIN scores is evident in the box plot in Fig. 5c, which reflects the bimodal distribution of M1 CIN scores observed in the heat map (Fig. 5a).

To identify genomic evidence of CIN in M1 cases, we sampled the same tumor regions for DNA extraction that were previously selected to generate transcriptome data in M0-poly and M0-NM cases. Since there is limited tissue in PNBX, only 24 cases yielded sufficient quality/quantity of DNA for CNA evaluation. However, a significant increase (p = 0.038) in copy number alterations in M1-poly versus M0-NM cases was observed in this small sample (Fig. 5d), consistent with heightened frequencies of amplifications and deletions associated with CIN in TCGA (Fig. 2). Gain of MYC and loss of RB1 and SIAH3 were also identified, consistent with previous studies of genetic alterations associated with poor PC prognosis (Additional File 1, Supplemental Fig. 6A and B) [42‐44].

The bimodal distribution of M1 cases when organized by CIN70 score suggests both CIN-dependent and CIN-independent gene associations and processes linked to PC lethality. Consequently, we evaluated DEGs and biological processes associated with CIN70-Low versus CIN70-High, as well as PC-CIN-Low versus PC-CIN-High cases from the PNBX cohort. Heat maps of the DEGs based upon these different gene expression signatures is displayed in Additional File 1, Supplemental Fig. 7A and B). Enriched biological processes associated with CIN70 and PC-CIN scores (low versus high) are also displayed (Additional File 1, Supplemental Fig. 7C and D). Distinct biological processes appear to be active in CIN-high versus CIN-low tumors. As expected, CIN-high tumors involve processes associated with cell cycle, mitosis, and chromosome segregation. In contrast, the top processes associated with CIN-low tumors involve developmental signatures, specifically those related to vascular and urogenital system development, as well as muscle contraction. In a subgroup analysis of clinical, pathological, and outcomes variables associated with CIN- high versus CIN-low metastatic tumors, there were no significant differences found in de novo metastatic cases. When all metastatic cases were included (i.e., those that were initially diagnosed as M0, but then progressed with metastasis), CIN-high cases were more likely to have pathological N1 stage and demonstrated significantly poorer outcomes than CIN-low cases (Supplemental Tables 4 &5).

Additional analysis of previously described CIN genes and drivers was also performed. During chromosome segregation, sister chromatids are separated by a kinetochore mediated attachment to spindle microtubules [45]. The microtubules are nucleated from centromeres, which require the highly evolutionary conserved OIP5/MIS18β for proper assembly [46]. Disruption of kinetochore and centrosome dynamics are components of neoplastic transformation, and, similar to aneuploidy, centrosome amplification is another hallmark of cancer [47]. Although there was no clear evidence of specific functional group dysregulation among nine mechanistic subgroups of CIN70 genes that we annotated (chromosomal separation, condensation, cyclins/cell cycle, DNA damage repair (DDR), kinetochores, spindle-related, spindle related/centrosomes, and spindle related/cyclins), we did find interesting expression differences that connect CIN and metastatic progression. KIF20A is one of the leading edge genes found on GSEA of M1-poly versus M0-NM DEGs (Fig. 4a) and is homologous to KIF2B, a protein that directly promotes tumor metastasis in cell line models of CIN [15]. Both KIF20A and KIF2B are significantly overexpressed in PC-CIN-high cases relative to (Additional File 1, Supplemental Fig. 8A).

A recent analysis of highly aneuploidy breast cancers in TCGA found overexpression of three transcriptional regulators, E2F1, MYBL2, and FOXM1 [13]. Overexpression of these genes in non-transformed Xenopus embryos was sufficient to significantly increase the rate of chromosomal missegregation and initiate aneuploidy. Evaluation of expression of these transcription factors in CIN70-high versus CIN70-low PNBX demonstrated significantly elevated expression in CIN70-high cases (Additional File 1, Supplemental Fig. 8B). In addition, 6/7 PC-CIN genes (CEP55, UBE2C, MELK, TPX2, PTTG1, and CDCA3) were also found to be among the top DEGs identified through comparison of high aneuploidy versus low aneuploidy breast tumors in TCGA. These results suggest that the same drivers and effectors are involved across tumor types.

Next, we questioned whether the PC-CIN signature genes were associated with disease progression in presumed localized (M0) cases. PC-CIN score separated high- and low-risk BCR groups from two independent PC cohorts (Fig. 6a) [22, 48]. We also tested the ability of PC-CIN to separate cases based upon M-stage equally well in subcohorts of African-American (AA) and European-American (EA) PNBX (n = 121). Clinical characteristics of patients included in this expanded PNBX cohort are shown in Additional File 1, Supplemental Table 2. PC-CIN was associated with metastatic progression in cases stratified by race with AUC of 0.78 for AA men and 0.80 for EA men (Fig. 6b). Both AA and EA men displayed significantly higher PC-CIN in M1 PNBX compared to M0 PNBX (P = 1.06e-04 and 3.11e-04, respectively, Fig. 6c). In both racial groups, univariate analyses demonstrated that PC-CIN high cases were significantly more likely to be classified as clinical stage M1, progress to CRPC, and die from PC (Additional File 1, Supplemental Table 3). In order to evaluate the impact of PC-CIN score in the context of other clinicopathological variables, multivariate logistic regression analysis was performed. After controlling for age, race, PSA and Gleason sum, PC-CIN-high is significantly associated with higher odds of M1 stage, CRPC, PC-death and all-cause mortality in multivariate analysis (OR 10.84, 16.13, 6.26 and 6.00, respectively; all p < 0.001, Table 1).