Cribriform and intraductal prostate cancer are associated with increased genomic instability and distinct genomic alterations

Via online access (http://​cancer.​digitalslidearch​ive.​net) and mScope Portal (Aurora Interactive, Montréal, Canada) three investigators with expertise in urogenital pathology (C.K., Th.v.d.K., and G.v.L.) reviewed available whole-slide images of frozen sections of both TCGA (n = 260) and CPC-GENE (n = 199) cohorts. Both cohorts contained radical prostatectomy specimens without prior hormonal or radiation therapy. Each slide was reviewed for GS, tumour percentage and percentage CR/IDC. Percentage CR/IDC was defined as estimated number of CR/IDC tumour cells divided by the total number of cells present in the tissue slice. Since invasive cribriform and IDC-P were morphologically indistinguishable, they were not scored individually [13].

All statistical analyses were performed in the statistical programming language R v3.2.1 and all genomic coordinates in this manuscript are based on the latest hg19 genome build. Gene-wise log₂ ratios for revised TCGA PRAD samples (based on Affymetrix SNP 6.0 arrays) were retrieved via the TCGA-Assembler R-package [30]. To obtain discrete values, gains or deletions of genetic regions were called if a sample’s copy number exceeded the threshold of ±log₂(1.5/2). Similarly, a gene-by-sample matrix was obtained for all revised CPC-GENE samples based on Affymetrix OncoScan arrays as described in [17]. Percent genome altered (PGA) was calculated for both the whole genome (excluding chrX and chrY) as described in [17] and separately for individual chromosome arms. For chromosome arms, separate PGAs for amplifications and deletions were obtained by dividing the number of bases affected by a deletion/amplification by the number of bases of the respective chromosome arm, taking into account only one DNA strand as PGA does not account for the strand of CNAs. For all values, a Wilcoxon-Mann-Whitney test was performed to test for significant differences between GS categories.

For identifying CR/IDC-associated events, the TCGA cohort was used as discovery set and the CPC-GENE cohort was used for validation. We initially used all CR/IDC positive samples for our analyses, but subsequently limited the CR/IDC group to cases with at least 30% to account for possible signal losses due to dilution effects caused by non-CR/IDC tissue without CNAs. This dilution effect can be envisioned assuming that CNAs of interest are CR/IDC-associated and corresponding signals therefore mainly originate from the CR/IDC compartment of the tumour. Surrounding non-CR/IDC tissue hence does not harbor these CNAs and only contributes to background signal leading to a reduced signal-to-noise ratio when trying to detect the CNAs in a mixture of both tissues. Prior to analysis, duplicated gene names, known read-throughs, genes on non-random/haplotype chromosomes, as well as genes in pseudoautosomal regions and with missing data were removed. After these filtering steps, 22,350 and 22,420 genes remained for analysis of the TCGA and CPC-GENE cohort, respectively. Next, adjacent genes exhibiting the same CNA profiles were grouped into regions to further reduce the number of tests. Boschloo’s exact test (one-sided, R-package ‘Exact’) was applied to regions with CNAs in at least 10% of all samples to identify events that occurred significantly more often in samples with CR/IDC. Multiple testing correction was performed via false discovery rate (FDR) and regions with a q-value below 0.05 were considered significant. To integrate both cohorts, all genes in regions that were identified as significant in the TCGA cohort were tested in the CPC-GENE cohort. Genes with a q-value below 0.1 were considered validated. A logistic regression was used to assess which individual deletion or amplification events were predictive for CR/IDC status while accounting for PGA and GS as confounding factors. To account for correlations between PGA and individual CNAs, PGA was re-calculated for each event by excluding the chromosome the particular event was located on. Visualization of results was done with BoutrosLab.plotting.general R-package (v5.6.10; P’ng et al. in review).

To quantify ERG expression in the TCGA cohort, RSEM ‘scaled estimates’ were obtained via TCGA-Assembler and multiplied by 10⁶ to convert them to transcripts per million (TPM). Subsequently a log₁₀ transformation was applied and UCSC transcript uc002yxa.2 was used to estimate ERG expression. Deletion events located between TMPRSS2 and ERG were determined by combining deletions of the genes ETS2, BACE2, BRWD1, PSMG1 and HMGN1. For the CPC-GENE cohort, scores for chromothripsis and kataegic regions were computed using the ShatterProof [31] and SeqKat (Fraser et al. Nature, in press) algorithms. The maximum values for each sample were used for comparison (Wilcoxon-Mann-Whitney test) to ascertain that despite their rare occurrence, any presence of these phenomena in the CPC-GENE samples could be detected and tested for association with CR/IDC.

Automated and curated somatic mutation calls for exome sequencing data from TCGA PRAD samples were obtained via the TCGA Data Portal (https://​tcga-data.​nci.​nih.​gov/​). Functional events were summarized patient-wise for each gene (i.e. multiple mutations in one gene were only counted once per patient, excluding categories ‘Silent’ and ‘RNA’). In addition, non-recurrent events and events that occurred in less than 5% of all tested samples were excluded from further analysis; all remaining gene mutations were tested for significant enrichment in CR/IDC positive samples using Boschloo’s exact test (one-sided, R-package ‘Exact’). CPC-GENE whole genome sequencing-derived SNVs (Fraser et al. Nature, in press) were filtered to only include functional mutations located in exonic regions and then processed as described above.

Patient characteristics of both TCGA (n = 260) and CPC-GENE (n = 199) cohorts are listed in Table 1. The TCGA cohort included more patients with adverse characteristics than the CPC-GENE cohort, having higher PSA levels (Wilcoxon rank sum test, p = 2.2·10⁻¹⁶), GS (Pearson’s χ2 test, p = 4.0·10⁻⁵) and pT stage (Pearson’s χ2 test, p = 3.1·10⁻⁹), which can be explained by the specific inclusion of clinically intermediate-risk disease in the latter cohort. Moreover, tumour cellularity was higher in TCGA than CPC-GENE (Additional file 1: Figure S1). Representative prostate cancer samples of GS 6 and GS ≥ 7 are depicted in Fig. 1.

To assess whether CR/IDC was associated with genomic instability, we calculated PGA for all patients and used a Wilcoxon-test to identify significant differences [17, 26]. PGA was 3 fold (p = 1.6·10⁻⁴) higher in men with CR/IDC as compared to men without (Fig. 2). Exclusion of men with GS 6, who generally lack CR/IDC growth, yielded similar results with 2.2 fold (p = 3·10⁻⁴) PGA increase in cases containing CR/IDC. Subgroup analysis revealed that PGA was significantly higher in samples with CR/IDC in GS 4 + 3 = 7 (2.2 fold; p = 5.3·10⁻³), but not in GS 3 + 4 = 7 (2.1 fold; p = 0.19), GS 8 (5.1 fold; p = 0.57) and GS 9–10 (1.7 fold; p = 0.10). Moreover, PGA scores did not differ significantly between GS 3 + 4 = 7 without CR/IDC pattern and GS 6 (1.2 fold; p = 0.51). Validation within the CPC-GENE cohort revealed overall 1.7 fold higher PGA of CR/IDC positive men with GS ≥ 3 + 4 = 7 (p = 4·10⁻³). Subgroup analysis showed 1.3 fold (p = 0.02) higher PGA in GS 3 + 4 = 7 cases with CR/IDC as compared to those without. PGA scores were significantly lower in GS 6 as compared to GS 3 + 4 = 7 with CR/IDC (2.2 fold; p = 4.7·10⁻⁷) than those without CR/IDC (1.6 fold; p = 0.07). Since 32 out of 35 CPC-GENE patients with GS ≥ 4 + 3 = 7 had CR/IDC, statistical analysis in respective subgroups lacked statistical power.

To determine whether genomic instability in CR/IDC was a global phenomenon or affected specific genomic regions, we computed PGA for individual chromosome arms utilizing deletion and amplification events independently. We found that deletions were mostly present on chromosome arms 1p, 4p, 4q, 5q, 7q, 8p, 10p, 10q, 12p, 13q, 16q, 17p, 18q and 21q in samples with CR/IDC (p < 0.05, Additional file 1: Figs. S2 and S3; Additional file 2: Table S1), while amplifications were found on chromosome 4q, 8p, 8q, 9p, 14q and 18p. Several of these chromosome arms have been linked to advanced prostate cancer [21, 32‐35]. Increased PGA for chromosome 4p, 8p, 10q, 12p and 16q deletions were also present in the CPC-GENE cohort (p < 0.05, Additional file 1: Figs. S4 and S5; Additional file 2: Table S1).

To identify somatic CNAs associated with CR/IDC, we applied Boschloo’s exact test, independently for each gene locus in GS ≥ 3 + 4 = 7 samples. We found 592 gene deletions and 366 amplifications significantly associated with CR/IDC (q < 0.05). These events clustered in specific chromosomal regions known to be associated with aggressive disease such as deletions of 8p (PPP2R2A, NKX3–1) [36‐38], 16q22 (CDH1) [39], 16q23 (BCAR1, CTRB1, CTRB2, WWOX and MAF) [15, 40, 41], 16q24 [42], 10q23 (PTEN) [43, 44], 17p13 and 18q21 (CCBE1) [45] as well as amplification of 8q24 (MYC and LY6 family members [15, 46, 47], Fig. 3 and Additional file 3: Table S2).

Since it was unclear whether genomic alterations occurred specifically in CR/IDC structures or also in non-cribriform prostate cancer glands adjacent to CR/IDC, we excluded samples with <30% CR/IDC growth pattern. Comparing GS ≥ 3 + 4 = 7 men with ≥30% CR/IDC (n = 44) to those without (n = 84) resulted in a total of 1299 significant deletions and 369 amplifications. Additional deletions in cases with ≥30% CR/IDC included the “Down syndrome critical region” located between ERG and TMPRSS2 on 21q22 [48], 16q22 (CTCF) [49], 13q14 (RB1) [50, 51], 17p13 (TP53) [52], and parts of 6q [53, 54] (Additional file 4: Table S3). Although genetic deletions of genes located between the TMPRSS2 promoter and ERG occurred more frequently in CR/IDC cases, we were unable to find a significant difference in ERG mRNA expression (Additional file 1: Figure S6). This paradoxical finding might be explained by relatively more frequent genomic translocation than deletion mechanism for TMPRSS2:ERG corresponding to lower genomic instability in cases without CR/IDC [55].

A trend towards lower q-values was observed when excluding tumours with <30% CR/IDC pattern suggesting that signal strength from CR/IDC specific events was diluted in cases with low CR/IDC quantity. Subsequent analyses were all performed using CR/IDC samples with at least 30% cribriform architecture. In total 474 deleted and 328 amplified genes were validated in the CPC-GENE cohort (q < 0.1), located on chromosomes 8p, 10q23, 13q22, 16q23–24, 17p13, 21q22, as well as 8q24, respectively (Additional file 5: Table S4 and Additional file 6: Figure S7). We noticed that q-values were generally lower in TCGA as compared to CPC-GENE, regardless of whether a threshold on CR/IDC was applied or not, indicating relatively lower statistical power of the latter cohort.

Since genomic instability and GS might act as confounding factors in assessing CNA events, we performed logistic regression analysis correcting for GS and PGA based on the 1668 previously identified events. A total of 779 gene deletions and 317 amplifications were independently associated with CR/IDC (q < 0.1, Additional file 7: Table S5). Deletions were mostly located on 8p21–23, 13q14, 16q21–24 as well as 18q21–23, but also included the genomic loci containing PTEN (10q23) [56], RYBP/FOXP1 (3p13) [16] and CASP8AP2 (6q15) [57]. The PPP2R2A/BNIP3L/PNMA2 locus (8p21) [36] featured the lowest q-value for deletions (p = 0.00018, q = 0.02, OR = 10.2, 3.24–38), while the MAFA/PTP4A3 locus on 8q24 did for amplifications (p = 0.007, q = 0.08, OR = 7.77, 1.98–41.95) [58, 59]. For CPC-GENE, logistic regression did not yield significant results after correcting for multiple comparisons, which can be attributed to lower statistical power and significant differences in pathological features.

To identify genes affected by functional SNVs we used TCGA exome sequencing data (https://​tcga-data.​nci.​nih.​gov/​) of samples with GS ≥ 7, and compared 88 samples with ≥30% CR/IDC against 143 without. Filtering for genes that harboured SNVs in at least 5% of all samples, FOXA1 (15% versus 5%; p = 0.007), TP53 and SPOP (both 19% versus 10%; p = 0.035) showed significantly higher mutation rates in cases with CR/IDC compared to those without (Boschloo’s exact test). Although SNV data were available for CPC-GENE samples, the number of cases, i.e. 8 with and 30 without CR/IDC was too low for statistical analysis. We did not find significant differences in overall frequency or total number of affected genes with functional SNVs (data not shown), indicating that SNVs are unlikely to be driver events for CR/IDC growth.

Finally, we investigated whether recently discovered DNA repair-related phenomena were linked to CR/IDC [60, 61]. We utilized available computational scores for kataegis, a pattern of localized hypermutation, and chromothripsis, a catastrophic event during which single chromosome arms or entire chromosomes are rearranged and/or lost. No statistically significant differences could be identified between cases with and without CR/IDC albeit sample numbers were low (data not shown).