Genomic Hallmarks and Structural Variation in Metastatic Prostate Cancer

SUMMARY

While mutations affecting protein-coding regions have been examined across many cancers, structural variants at the genome-wide level are still poorly defined. Through integrative deep whole genome and transcriptome analysis of 101 castration-resistant prostate cancer metastases (109X tumor / 38X normal coverage), we identified structural variants altering critical regulators of tumorigenesis and progression not detectable by exome approaches. Notably, we observed amplification of an intergenic enhancer region 624 kilobases upstream of the androgen receptor (AR) in 81% of patients, correlating with increased AR expression. Tandem duplication hotspots also occur near MYC, in lncRNAs associated with post-translational MYC regulation. Classes of structural variations were linked to distinct DNA repair deficiencies, suggesting their etiology, including associations of CDK12 mutation with tandem duplications, TP53 inactivation with inverted rearrangements and chromothripsis, and BRCA2 inactivation with deletions. Together, these observations provide a comprehensive view of how structural variations affect critical regulators in metastatic prostate cancer.

Graphical Abstract

In brief

Integrative whole genome and transcriptome sequencing provides a comprehensive view of structural variations that affect major regulators in prostate cancer and would escape detection by exome-based approaches.

INTRODUCTION

Prostate cancer represents a common and clinically heterogeneous disease entity. While over 160,000 American men are diagnosed with prostate cancer each year, less than 20 percent of patients will experience progression to the lethal form of the disease, termed metastatic castration-resistant prostate cancer (mCRPC) (Siegel et al., 2018). A major barrier to studying mCRPC has been the difficulty in obtaining tumor samples, as clinical biopsies of metastatic lesions are not routinely performed. mCRPC has recently been evaluated by targeted or whole exome sequencing (Armenia et al., 2018; Beltran et al., 2013; Grasso et al., 2012; Robinson et al., 2015; Zehir et al., 2017). These studies identified alterations in pathways involving androgen signaling, DNA repair, and phosphoinositide 3-kinase (PI3K) signaling, as well as recurrent mutations in genes such as SPOP, FOXA1, and IDH1. However, the exome represents less than 2% of the genome, and outside of small case series (Gundem et al., 2015; Wedge et al., 2018), the complete genomic landscape of mCRPC remains largely unexplored.

Genomic structural variants (SVs) include genomic deletions, insertions, tandem duplications, inversion rearrangements, and inter-chromosomal translocations. SVs are prevalent in prostate cancer, with gene fusions involving the E26 transformation-specific (ETS) family of transcription factors identified in 40–60% of cases (Maher et al., 2009; Tomlins et al., 2007; Tomlins et al., 2005). A recent study in localized prostate cancer demonstrated clusters of genomic rearrangements each occurring in 5–6% of samples (Fraser et al., 2017). In addition, previous studies have demonstrated that SVs may define subtypes of ovarian, pancreatic, and breast cancers (Nik-Zainal et al., 2016; Patch et al., 2015; Waddell et al., 2015; Wang et al., 2017). Of note, the majority of SVs involve intergenic or intronic noncoding regions of the genome and are not captured by exome sequencing or transcriptome analysis. A key advantage over exome sequencing is that whole genome sequencing (WGS) allows the identification of SVs that alter the activity of key driver genes, tumor suppressors, and regulatory elements.

To comprehensively investigate the genomic drivers of mCRPC, we interrogated the whole genomes and transcriptomes of mCRPC samples from over 100 patients at a mean depth of 109X in tumors, a depth 23 times greater than that achieved in previous large WGS studies in cancer. Deep sequencing of a large patient cohort permitted us to discover novel recurrent SVs and define the prevalence of these variations in mCRPC. We discovered previously unidentified recurrent SVs modulating tumor suppressors or oncogenes, identified new rearrangements coupling noncoding genes to known cancer drivers, and uncovered novel global associations between DNA repair alterations and SVs.

RESULTS

A multi-institutional consortium conducted a prospective IRB-approved study (NCT02432001) that obtained and profiled metastatic tumor biopsies from prostate cancer patients with castration-resistant disease (Aggarwal et al., 2016). Image-guided core biopsies were obtained (Holmes et al., 2017) and fresh-frozen. Tumor tissue was centrally processed and banked. Laser capture microdissection was used to isolate samples enriched for cancer, and sequencing of RNA was performed. Whole genome DNA sequencing was performed from frozen sections for tumor and from peripheral blood for matched normal samples, obtaining a mean depth of 109X in tumor and 38X in normal samples (Figure S1). Paired end mRNA libraries were sequenced to a median depth of 114M paired reads. This report includes results from 101 patients, including mCRPC lesions from bone (N=42), lymph node (N=40), liver (N=11) or other soft tissue sites (N=8) (clinical summary in Table 1, sample-level features related to sequencing, molecular analysis, and biopsy site in Table S1). Of these patients, 64% had received second-generation anti-androgen therapy (abiraterone: 47%, enzalutamide: 37%, both: 20%).

Table 1:

clinical characteristics of the patient cohort

category	value
median age (range)	71 (45–90)
race/ethnicity
white	85
black	5
asian	4
unknown	7
Gleason grade at diagnosis
6	11
7	28
≥8	52
unknown	10
site of biopsy
bone	43
lymph node	39
liver	11
other soft tissue	8
prior therapy
Abiraterone	27
Enzalutamide	17
Both	20
Neither	37
visceral metastases
yes	31
no	70
median lab values at biopsy
PSA, ng/mL (range)	65.1 (0.4–1874.5)
Alkaline phosphatase, IU/L (range)	92 (49–1506)
Lactate dehydrogenase, IU/L (range)	187 (31–856)
Hemoglobin, g/dL (range)	12.8 (8.0–15.7)

Structural variations disrupt key driver genes

The frequency of genomic copy number alterations in our mCRPC tumors was consistent with previous exome sequencing reports (Armenia et al., 2018; Beltran et al., 2013; Grasso et al., 2012; Robinson et al., 2015) (Figure 1A, Figure S2A). The percent of the genome altered in each sample ranged between 7% and 47% (median 23%; Table S1). The median mutation frequency was 4.1 mutations/Mb., slightly lower than the 4.4 mutations/Mb. reported previously in mCRPC (Robinson et al., 2015), but greater than the 0.53 mutations/Mb. reported in primary prostate cancer (Fraser et al., 2017). Approximately 40% of tumors were triploid (Figure S2B, Table S1), which was associated with more translocations and mutations overall (P < 0.007, Figure S2C, S2D)

Figure 1.

Structural Variants Disrupt Tumor Suppressors and Activate Oncogenes

(A) SV and copy number frequency plotted on scaled chromosomes. Wider green/blue bars indicate more frequent copy gain/loss. Darker black bars indicate more frequent SV.

(B) Top: expression levels of PTEN, TP53, RB1, CDKN1B, and CHD1 in individual samples reported as (log[1+(TPM × 10^{^6})]). Bottom: somatic events affecting each sample. Right: box and whisker plots showing expression for samples with 0, 1, or 2 alleles affected; horizontal bar indicates median. Each gene was sorted independently by expression level. See also Figure S3.

(C) Schematic diagrams of ETS family fusions indicating previously observed and novel partners. See also Figure S3.

(D) Schematic diagram of ETV1 activation via RP11-35609.1 fusions. See also Figure S4.

(E) Schematic diagrams of oncogene fusions showing previously observed and novel partners.

See also Figure S2 and Tables S1, S2, S3, and S5.

We systematically identified loci most frequently affected by structural variations by counting SVs within one megabase windows genome-wide (SV per window 9.6 ± 5.1; mean ± SD, listed in Table S2). The frequency of SVs is plotted in concert with copy number alteration frequencies in Figure 1A. The loci most frequently affected by SV (>3 SD from mean) contained key drivers of prostate cancer, underscoring the importance of structural variation in this disease. This included AR, the transmembrane serine protease 2 (TMPRSS2) and ETS transcription factor (ERG) genes that produce the TMPRSS2/ERG fusion, the oncogene MYC, Forkhead Box protein A1 (FOXA1), and phosphatase and tensin homolog (PTEN). This analysis also identified clusters of deletions affecting genes at fragile sites previously identified in more than one cancer type (Bignell et al., 2010; Glover et al., 2017).

An integrated analysis of SVs and mRNA expression levels was then used to define cases where SVs were predicted to inactivate tumor suppressors. PTEN was affected by biallelic alterations in 36% of tumors and monoallelic alterations in 26% of tumors (Figure 1B). The PTEN sequence or promoter was frequently interrupted by a translocation (7% of cases) or inverted rearrangement (5% of cases, Figure S3A). SV were essential to assigning biallelic PTEN alteration status in 8% of cases, and mono-allelic PTEN alteration status in 5% of cases (Figure 1B, left). TP53 was affected by biallelic somatic alterations in 46% of tumors and monoallelic alterations in 30% of tumors, with 11% of the biallelic assignments due to SV gene disruption. SVs also contributed to biallelic inactivation of RB1 (12% biallelic, 3% by SV), CDKN1B (7% biallelic, 1% by SV), and CHD1 (7% biallelic, 2% by SV) (Figure 1B). There was a significant association between the number of inactivated alleles and mRNA levels of PTEN, TP53, CDKN1B, RB1, and CHD1 (Figure 1B, right), suggesting monoallelic alterations impacted expression levels of these genes.

Novel gene fusions predicted to activate oncogenes

We then determined cases where structural variants were predicted to activate driver genes by integrating SV data, mRNA expression levels, and predicted mRNA fusions. A majority of prostate cancers harbor fusions from the juxtaposition of the 5’ regulatory region of the androgen-responsive gene TMPRSS2 upstream of ERG (Tomlins et al., 2005). We observed mutually exclusive fusions activating the ETS family members ERG, ETV1, ETV4, and ETV5 in 59% of our cohort (Figure 1C, fusions listed in Table S3). In four cases, an ETS family member fused to a gene not previously reported in mCRPC, including ETV1 fusions driven by the solute carrier SLC30A4 and ETV4 fusions driven by Transmembrane and Coiled-Coil Domain Family 2 (TMCC2), Clathrin Heavy Chain (CLTC), and Cell Division Cycle 6 (CDC6). We also identified novel fusions between coding and non-coding genes, exemplified by SCHLAP1, a lncRNA highly enriched in a subset of aggressive prostate cancers (Prensner et al., 2013). The PI3K pathway member PIK3CA was expressed at very low levels except for a single sample bearing a translocation that placed the first exon of SCHLAP1 immediately downstream of the PIK3CA 5’ UTR, resulting in the overexpression of a full-length PIK3CA transcript (Figure S3B). In two other cases, ETV1 was translocated to chromosome 14 between FOXA1 and Mirror-Image Polydactyly 1 (MIPOL1). The lncRNA RP11–356O9.1 (also annotated as AL121790.1) lies in this region. Previously published data showed that in normal tissues, RP11–356O9.1 is expressed exclusively in prostate (Figure S4). In these two cases, the first exon of RP11–356O9.1 was fused to exon 4 or exon 5 of ETV1 (Figure 1D). Fusions between ETV1 and this region have been previously reported in the prostate cancer cell line MDA-PCa 2B (Tomlins et al., 2007) and in a single patient sample (Abeshouse et al., 2015).

Multiple low-frequency gene fusions involving oncogenes, including AXL, BRAF, and MYC, were also noted (Figure 1E, Figure S3B). A gene fusion joined prostatic acid phosphatase (ACPP) residue 380 (NM_001009) to the transmembrane receptor tyrosine kinase AXL at residue 429 (NM_001699), producing an in-frame transcript. Review of an independent cohort of patients with high risk primary prostate cancer identified a similar ACPP-AXL fusion, demonstrating these fusions are a repeated finding (Figure S3C). In a case lacking high level MYC DNA copy number amplification, ACPP was fused to MYC within 150 nt of the MYC 5’ untranslated region, originating within the second and third ACPP exons. Collectively, these novel, low frequency gene fusions could represent therapeutic targets in mCRPC.

Duplications target AR, MYC, and FOXA1

Genomic duplication events are a mechanism of genome evolution (Ohno, 1970) and are known to alter specific drivers important in cancer, such as FLT3 in acute myeloid leukemia and BRAF in pilocytic astrocytoma (Jones et al., 2008; Nakao et al., 1996). Unbiased analysis identified a region approximately 624 Kb. upstream of AR as the most frequent site of structural variation in mCRPC (Table S2). AR amplification occurred in 70% of cases, and was associated with significantly elevated AR mRNA expression (P = 9 × 10⁻⁸, Figure 2A top, Figure 2B). Our result is consistent with earlier findings that AR amplification is rare in primary prostate cancer (Abeshouse et al., 2015) but common in mCRPC (Robinson et al., 2015), and is a major mechanism of resistance to androgen deprivation therapy (Visakorpi et al., 1995). The region of peak amplification upstream of AR at 66.94 Mb was amplified in 81% of cases, 11% more frequently than AR itself (Figure 2A, middle). Tumors frequently amplified both AR and the upstream peak (68 cases), but in 13 cases the upstream peak alone was amplified (Figure 2B). DNA copy gain at the upstream peak in cases that lacked AR amplification was significantly associated with elevated AR expression (P = 0.003, Figure 2B), indicating that amplification of the upstream peak was independently associated with AR expression levels. Cases with amplification of both the upstream peak and AR had significantly higher expression than cases where only the upstream peak was amplified (P = 0.01, Figure 2B), consistent with additive effects.

Figure 2.

Tandem Duplication Target Enhancers near AR, MYC, and FOXA1

(A) Aligned tracks showing the DNA amplification frequency (top), tandem duplication frequency (middle), tandem duplication bounds (middle), and H3K27ac average read coverage (bottom, from Kron et al., 2017) at the AR locus.

(B) Box and whisker plot showing AR expression in the presence/absence of DNA amplification at AR or at the peak.

(C) Samples with tandem duplication of the peak in (A) but not AR (red) more frequently had AR unamplified or amplified at low levels.

(D and E) aligned tracks showing tandem duplications near MYC (D) and FOXA1 (E) as in (A).

See also Table S4.

Tandem duplications at the upstream peak corresponding to copy number gain break points were observed in 36% of all cases, and in 44% of the 81 cases bearing copy gain at this region. Focal tandem duplication of the upstream peak region was almost exclusive to patients lacking or with low AR amplification (P < 0.0007, hypergeometric test, Figure 2C), consistent with tandem duplication at the peak being a sufficient alternative to AR amplification. The presence of amplification at this peak was not associated with previous treatment with the second-line hormone therapies abiraterone or enzalutamide (Table S4). We assessed the frequency of H3K27ac occupancy within the upstream peak, as H2K27ac enrichment is associated with potential enhancer activity (Heintzman et al., 2009). Previously published data from 19 primary prostate tumors revealed that the minimally targeted region at the upstream peak, was enriched for H3K27ac histone modifications (Kron et al., 2017) (Figure 2A, middle & bottom). Collectively, these data support the detection of an enhancer, amplified in 81% of castration-resistant metastatic patients, that can act independently of AR locus amplification to increase expression of AR in response to first-line ADT.

Intergenic regions near MYC at 8q24 and FOXA1 at 14q13.3 were also frequent targets of SVs (Table S2). We observed distinct tandem duplication peaks 700 and 300 Kb. upstream of MYC, with duplication frequencies of 25% and 23% respectively (Figure 2D, top, middle). The farther region included three long non-coding RNAs: prostate cancer associated transcript -1 and -2 (PCAT-1, PCAT-2), and prostate cancer associated non-coding RNA 1 (PRNCR1). The degree of MYC copy number amplification was modestly associated with MYC mRNA expression levels (rho = 0.28, P = 0.005). Although PRNCR1 is unlikely to be implicated in mCRPC pathogenesis (Prensner et al., 2014), PCAT-1 has been shown to upregulate cMyc protein levels post-translationally (Prensner et al. 2014b). The nearer region included additional non-coding genes, as well as the rs6983267 and rs1447295 germline variants associated with prostate cancer risk (Amundadottir et al., 2006; Yeager et al., 2007). Tandem duplications overlapping FOXA1 and/or the adjacent gene mirror-image polydactyl 1 (MIPOL1) were present in 14% of samples (Figure 2E). These events were less frequent than AR or MYC events described above, precluding nomination of a candidate local peak, but several sites in this region had H3K27ac enrichment (Figure 2E). Three of the fourteen samples bearing tandem duplications in this region also bore FOXA1 mutations. Observations of ETV1 translocations into this region by us (Figure 1D) and others (Abeshouse et al., 2015; Tomlins et al., 2007) suggest that SVs at this locus plays a role in prostate cancer. These observations collectively demonstrate that unbiased analysis of tandem duplications by whole genome sequencing identifies loci that are selected for amplification near driver genes such AR, MYC, and FOXA1 in metastatic prostate cancer, and that this selection potentially drives disease progression.

DNA repair defects drive SVs

To explore the etiology of SVs in prostate cancer, we identified alterations associated with SV frequency. The number of SVs identified in individual tumors ranged between 103 and 923 (337 ± 166, mean ± SD, Figure 3A). Deletion frequency was significantly higher in tumors with biallelic BRCA2 mutations (P = 4 × 10⁻⁶) (Figure 3B left, Figure 3C). Additionally, we observed that biallelic inactivation of CDK12 was associated with a significant increase in tandem duplications with a bimodal length distribution (Figure 3B center, Figure 3C, Figure S5) (P = 0.003). These results were consistent with results previously reported in ovarian cancer (Popova et al., 2016).

Figure 3.

DNA Repair Alterations Are Associated with Structural Variation Frequency

(A) Top: structural variant frequency by sample, sorted by deletion frequency. Bottom: presence of chromothripsis or biallelic inactivating alterations in BRCA2, CDK12, or TP53.

(B) Circos plots illustrating BRCA2 inactivation (left), CDK12 inactivation (center), and chromothripsis (right). Colors as in (A).

(C) Box and whiskers plots showing association between biallelic inactivating alterations in BRCA2, CDK12, or TP53 and the frequencies of deletions, tandem duplications, and inverted rearrangements respectively. See also Figure S5.

(D) Counts of inverted rearrangements and deletions per sample. Samples with biallelic BRCA2 loss drawn in blue, samples bearing chromothripsis drawn in orange.

(E) Box and whisker plots showing mutation frequency in the presence of biallelic loss of BRCA2 and chromothripsis.

See also Table S1.

We noted that the number of inverted rearrangements and deletions observed in each sample was significantly correlated (rho = 0.54, P < 4 × 10⁻¹⁰, Figure 3D). Tumors bearing large numbers of both deletions and inverted rearrangements had all undergone chromothripsis, the shattering and subsequent reconstruction of a single chromosome (Figure 3C, right; Figure 3D, orange points) (Fraser et al., 2017; Maher and Wilson, 2012; Stephens et al., 2011; Zack et al., 2013) (Figure 3B, right). We identified chromothripsis in 23% of mCRPC (Figure 3A, 3D, samples listed in Table S1), compared with 20% reported in non-indolent primary prostate tumors (Fraser et al., 2017). Biallelic TP53 inactivation was the event most significantly associated with elevated inverted rearrangement frequency (median 57 vs. 79 inversion rearrangements, P = 0.0004) and with the presence of chromothripsis (19 of the 23 cases with chromothripsis vs. 28 of the 78 cases lacking chromothripsis, P = 0.0004). No locus was preferentially targeted by chromothripsis, consistent with a stochastic process. No tumor with biallelic loss of BRCA2 also exhibited chromothripsis (Figure 3A, 3D). As observed in a previous pan-cancer analysis, chromothripsis was not associated with an elevated mutation frequency genome-wide (P > 0.05, Figure 3E) (Zack et al., 2013). In contrast, BRCA2 loss had the strongest statistical association with tumor mutational burden (median 7.0 vs. 4.0 mutations/Mb, P = 0.0002, Figure 3E).

Chromoplexy, a balanced interweaving of interchromosomal translocations, has been observed in prostate cancer (Baca et al., 2013). We identified chromoplexy in 50% of samples (Table S1). Of the 23 samples with chromothripsis, 12 (52%) also showed chromoplexy, as expected if there were neither positive nor negative enrichment for chromothripsis in samples that had undergone chromoplexy. The presence of somatic TP53 alterations was not associated with either translocation frequency or with the presence of chromoplexy. Our analysis therefore identified biallelic inactivation of CDK12, BRCA2, and TP53 as strongly linked to three forms of SV in mCRPC, with the link between TP53 inactivation and inversion rearrangements further linked to chromothripsis.

Mutational signatures of DNA damage

Cells bearing homologous recombination repair defects develop genomic scars (reviewed in (Lord and Ashworth, 2016)), including deletions with homology at both ends of the deleted region. These cells rely on microhomology-mediated end joining to repair double strand DNA breaks, also known as alternative nonhomologous end-joining (Davies et al., 2017; Nik-Zainal et al., 2012; Nik-Zainal et al., 2016; Tutt et al., 2001). Tumors bearing biallelic loss of BRCA2 had elevated levels of deletions with flanking microhomology (Figure 4A). Tumors with biallelic inactivation of CDK12 or ATM, or with monoallelic alterations in BRCA1 or BRCA2, lacked this phenotype, confirming previously published observations (Polak et al., 2017). We fitted published mutation signature profiles to somatic single nucleotide variations and performed de novo mutational profile signature analysis using non-negative matrix factorization (Alexandrov et al., 2013). A solution including eight de novo signatures provided the optimal balance between variance explained and parsimonious modeling. Signature de novo 8 was strongly associated with samples bearing biallelic BRCA2 inactivation (Figure 4A, Figure S6A) and closely resembled COSMIC signatures 3 and 8 (Figure 4A, Figure S6B), previously associated with defects in homologous recombination DNA repair (HRD) (Alexandrov et al., 2013; Nik-Zainal et al., 2016). COSMIC 3 signature fit was significantly elevated in samples bearing biallelic loss of BRCA2, consistent with previous reports in breast, ovarian, and prostate cancer (P = 4 × 10⁻⁷, Figure 4A, 4B).

Figure 4.

Mutational Signatures of DNA Damage in mCRPC

(A) From top to bottom: the frequency of deletions bearing two or more nucleotides of microhomology; fit of mutation signatures COSMIC 3 and 8 and de novo 8; alterations associated with DNA repair by homologous recombination. See also Figure S6.

(B) Box and whisker plots showing mutation frequency in samples bearing either biallelic loss of BRCA2 or compound BRCA1-BRCA2 heterozygosity, compared to samples lacking either of these alterations.

(C) Box and whisker plots showing COSMIC signature 3 fit in tumors bearing biallelic loss of BRCA2 and samples bearing compound BRCA1-BRCA2 heterozygosity.

A sample with heterozygous mutation of both BRCA1 and BRCA2 lacked an elevated microhomology deletion frequency, but nevertheless showed strong de novo 8 and COSMIC 3 signature scores. In all, 6% of cases harbored compound BRCA1/2 heterozygosity, either by single copy DNA loss (N = 5), or somatic mutation (N = 1). These samples had significantly elevated mutation frequency, statistically indistinguishable from that observed in BRCA2^−/− samples (Figure 4B). Compound heterozygous samples had COSMIC 3 signature scores intermediate between cases with biallelic BRCA2 inactivation and cases with one or zero BRCA1/2 alleles affected (Figure 4C), but the difference in signature fit was not statistically significant.

The other robust de novo signatures identified in this cohort recapitulated known signatures (Alexandrov et al., 2013). These included de novo 1, likely identical to COSMIC signature 1 associated with spontaneous deamination of 5-methylcytosine associated with age at tumor diagnosis (Alexandrov et al., 2015), and de novo 5, present in a hypermutated sample with deep deletion of mutS homolog 2 (MSH2) and MSH6, mismatch repair genes 300 Kb. apart on chromosome 2. This signature bore the strongest similarity to COSMIC 6 (associated with defective MMR) and COSMIC 9 (activation-induced deaminase activity during hypermutation). These data confirmed that DNA repair defects in mismatch repair and homologous recombination can produce genomic scars in metastatic prostate cancer, and showed that BRCA1/2 compound heterozygosity produces a mutational phenotype distinct from that of biallelic BRCA2 inactivation.

A landscape of mutations and structural alterations

Somatic alterations and structural variants for 44 key prostate cancer genes are shown in Figure 5, and listed in Table S5. The somatic mutation frequencies were consistent with previous reports in mCRPC (Armenia et al., 2018; Robinson et al., 2015). In total, 85% of mCRPC samples carried either pathogenic activating AR mutations, amplifications of AR, or putative AR enhancer region amplifications, an increase over the 63% of cases identified as carrying AR alterations in a benchmark exome study of comparable size (Robinson et al., 2015). ETS family genes were activated by fusions in 59% of cases. We observed MAPK driver mutations in HRAS (p.Q61K, 2%) and BRAF (p.G469A, 1%). Putative dominant negative SPOP mutations were present in 5% of cases (Barbieri et al., 2012; Blattner et al., 2017). ETS gene family activations were mutually exclusive with activating alterations in the RAS/MAPK pathway members (P = 0.01, Fisher’s exact test) and with inactivation of SPOP and CHD1 (Barbieri et al., 2012; Burkhardt et al., 2013; Huang et al., 2011). A single IDH1 mutation at the previously reported p.R132C hot spot was observed (Abeshouse et al., 2015). Additionally, mutually exclusive alterations affecting genes that modulate the AR pathway (FOXA1, NCOR1, NCOR2, and ASXL2) were present in 29% of cases. Alterations in WNT pathway members CTNNB1, APC, and ZNRF3 that were predicted to activate WNT signaling were mutually exclusive in all but one of the 17% of cases where they were present. Previously unreported inactivating events targeting HDAC4 were present in 6% of cases. No somatic alteration was significantly associated with tissue biopsy site after accounting for multiple testing correction. We searched for recurrent point mutations affecting the promoter, enhancer, and UTR regions of 574 known cancer driver genes (Table S5). This analysis identified 101 mutations of unknown significance; no variant was significantly associated with expression or structural variation phenotypes.

Figure 5.

Landscape of Somatic and Structural Alterations in mCRPC

Mutation frequency (top) and germline or somatic alterations in key genes where such alterations were predicted to be functionally meaningful. Alteration frequency shown at right.

See also Tables S3 and S5.

We next assessed the frequency of mutations in genes responsible for DNA damage repair. Inactivating germline alterations were present in the DNA repair genes (BRCA2 and ATM) in 4% of samples, a slightly lower frequency than the approximately 10% frequency observed in a large study of metastatic prostate tumors (Pritchard et al., 2016). Somatic alterations alone accounted for five of the eight cases of biallelic BRCA2 inactivation and all three tumors carrying biallelic CDK12 inactivation. Biallelic BRCA2, CDK12, and ATM inactivating mutations were mutually exclusive, and the total frequency of biallelic BRCA2, CDK12, and ATM inactivation was 15%. Two hypermutated samples were present, consistent with the reported 3% frequency of mismatch repair defects in mCRPC (Robinson et al., 2015). One hypermutated sample bore deep deletion in mutS homolog 2 (MSH2) and MSH6, mismatch repair genes 300 Kb. apart on chromosome 2, an alteration predicted to abrogate mismatch repair.

DISCUSSION

In contrast to previously published large-scale analyses of primary and metastatic prostate cancer that have largely focused on the coding genome (Abeshouse et al., 2015; Armenia et al., 2018; Barbieri et al., 2012; Beltran et al., 2013; Fraser et al., 2017; Robinson et al., 2015; Taylor et al., 2010; Wedge et al., 2018), we have performed whole genome analysis of metastases from 101 mCRPC patients at 109x depth of coverage in tumor samples. This coverage, two-fold deeper than previous efforts in this space (Wedge et al., 2018), and performed on a large patient cohort, has produced a unique resource for dissecting structural variation in metastatic prostate cancer samples. Our data emphasize that structural variations may inactivate tumor suppressors by disrupting the coding region of these genes (Patch et al., 2015; Waddell et al., 2015), whereas both fusions and alterations affecting intergenic regulatory elements appear to activate driver genes. Fusions driving proteins such as AXL or BRAF that can be targeted therapeutically may open directions for new treatments in mCRPC. We derived insight into the etiology of structural variation, associating BRCA2, CDK12, and TP53 with deletions, tandem duplications, and chromothripsis. Our novel observation that non-coding RNAs such as SCHLAP1 and RP11–356O9.1 drive oncogene expression highlights the under-explored role of non-coding genes in mCRPC and will serve as the foundation for further studies of the non-coding genome.

All of the men in this study had developed resistance following front-line treatment with androgen deprivation therapy (ADT). A key finding made possible by our integrated analysis of the whole genome and transcriptome across a large population of mCRPC patients is that amplification of a putative enhancer region 624 Kb. distant from AR was present in 81% of men, and 85% had either amplification or pathogenic activating AR mutation. Our data support the model that amplification at the putative enhancer locus results in increased AR expression. In 13% of men, putative enhancer amplification was present without alterations in AR itself. This finding suggests that DNA copy gain affecting this locus, commonly by tandem duplication, may be a frequent mechanism by which prostate tumor cells initially develop ADT resistance (Karantanos et al., 2013). Observations of tandem duplication at putative enhancers near AR, MYC, and FOXA1 underline the value of whole genome analysis, even in diseases where exome analysis has been performed in large cohorts of patients.

We observed chromothripsis in 23% of mCRPC patients and demonstrated that chromothripsis was significantly associated with TP53 alterations. This observation supports the proposed but unproven mechanistic association between TP53 alteration and chromothripsis (Rausch et al., 2012), reviewed in (Maher and Wilson, 2012). However, TP53 alterations cannot be the sole driver of chromothripsis, as chromothripsis is not widespread in other tumors with high rates of TP53 inactivation such as high grade serous ovarian carcinoma (Zack et al., 2013). In our study, chromothripsis was mutually exclusive with biallelic inactivation of BRCA2, inconsistent with a model where cells lacking the ability to perform homologous recombination repair of double-strand DNA breaks would be predisposed to chromothripsis.

Our study linked biallelic inactivation of BRCA2, but not ATM or CDK12 that manifest flanking microhomology (Figure 4A). It is not yet clear what combination of genotype and genomic data will best identify the patients who will benefit from PARP inhibitor therapy (Mateo et al., 2015). The 6% of samples with compound BRCA1/BRCA2 heterozygosity lacked deletions with flanking microhomology, but had significantly increased mutation rates not statistically distinguishable from biallelic BRCA2 tumors. Dissecting the functional consequences of these alterations will have implications for patient selection when considering treatment with a PARP inhibitor (Lord and Ashworth, 2016).

Our study demonstrates the utility of whole genome analysis across a clinically relevant metastatic tumor cohort, as our analysis led to multiple discoveries that eluded existing exome-centric genomic investigations in the advanced disease setting. We have provided the first landscape of structural variants in mCRPC, a substantial mutational class in this disease that will serve as a repository for other researchers to continue exploring their biological and clinical significance. Our data also provides the foundation for further dissection of the non-coding genome through complementary profiling efforts (e.g. epigenetics) and subsequent preclinical studies that may have translational impact in prostate cancer patients.

STAR METHODS

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Felix Feng (Felix.Feng@ucsf.edu).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Patient Cohort

Patient tissue samples were obtained through the Stand Up 2 Cancer/Prostate Cancer Foundation- funded West Coast Prostate Cancer Dream Team project, a multi-center study that acquired biopsies of metastases from men with mCRPC. All patients were male and ranged in age from 45–90 years when biopsied. See also Table 1 for additional clinical details. Samples were obtained by image-guided core needle biopsy of metastatic lesions in bone, soft tissue, or an organ. Fresh-frozen tissue and peripheral blood drawn at the time of biopsy was shipped to a central facility at UCSF for laser-capture microdissection and DNA and RNA extraction. Human studies were approved and overseen by the UCSF Institutional Review Board. All individuals provided written informed consent to obtain fresh tumor biopsies and to perform comprehensive molecular profiling of tumor and germline samples.

METHOD DETAILS

Sample Preparation and DNA Sequencing

Biopsies identified by histological assessment (H&E, serial section) to contain at least 50% tumor were selected for genomic DNA (gDNA) isolation through microdissection of frozen sections (200–500 μm total section depth, Qiagen QIAamp Fast DNA Tissue Kit, Cat. 51404). Matched normal gDNA was extracted from peripheral blood drawn at time of biopsy (Qiagen QIAamp DNA Blood Mini Kit, Cat. 51104). Tumor and Normal DNA were quantified prior to library construction using PicoGreen (Quant-iTTM PicoGreen dsDNA Reagent, ThermoFisher Scientific, Catalog #P11496). Quantifications were measured using a Spectromax Gemini XPS (Molecular Devices). PCR-free paired-end libraries were generated by automated liquid handlers using 500–1000 ng input gDNA and the Illumina DNA Sample Preparation HT Kit. Pre-fragmentation gDNA cleanup was performed using paramagnetic sample purification beads (Agencourt AMPure XP reagents, Beckman Coulter). Samples were fragmented, and libraries sizeselected following fragmentation and end-repair using paramagnetic sample purification beads to enrich for short insert sizes. Final libraries were quantified by qPCR and evaluated for quality using gel electrophoresis separation. DNA libraries were denatured, diluted and clustered onto patterned flow cells using the Illumina cBot system with Illumina HiSeq X HD Paired End Cluster Kit reagents. Clustered patterned flow cells were loaded onto HiSeq X instruments and sequenced on 151 bp paired-end, non- indexed runs on independent lanes, using HiSeq X HD SBS Kit reagents. Illumina HiSeq Control Software (HCS), and Real-Time Analysis (RTA) wren used with the HiSeq X sequencers for real-time image analysis, and base calling.

RNA Sequencing

Tumor cores were fresh-frozen in OCT for gene expression analysis. Laser capture microdissection was performed on frozen sections to enrich for tumor content (Spritzer et al., 2013). Total RNA was isolated (Agilent Absolutely RNA Nano Prep, Cat. 400753) and samples of sufficient quality (Agilent Bioanalyzer RNA 6000 Pico, Cat. 5067–1513) were amplified using NuGEN Ovation RNA-Seq System V2. cDNA fragmentation was performed on a Covaris M220 sonicator to 200bp. Libraries were generated using NuGEN Ovation Ultralow System V2 for Illumina sequencing. RNA-Seq for 88 samples was performed on an Illumina NextSeq500 (High Output 150 cycle V2 reagents, Cat. FC-404–2002) in 2x76bp paired-end runs; 13 additional samples were sequenced on an Illumina HiSeq2500 at 2x100bp (10), 2x101bp (2) and 2x50bp (1) paired-end runs.

Whole Genome Sequencing Data Analysis

Whole genome FASTQ files were uploaded to Illumina BaseSpace Sequence Hub (http://basespace.illumina.com). The Whole Genome Sequencing (WGS) app version 7.0.1 was used to coordinate sample alignment to the NCBI GRCh38 PAR-masked with decoys hs38d1 reference genome (hg38-decoy) and subsequent analytical steps. Reads were aligned against hg38-decoy using the Isaac aligner version 04.17.06.15 (Raczy et al., 2013). Germline mutation analysis was performed using Strelka version 2.8.0 (Saunders et al., 2012) filtered to require an assignment of PASS and snpEff version 4.3g (Cingolani et al., 2012b) labels of “pathogenic”, “splice_donor”, “splice_acceptor”, “stop_gain”, or “frameshift”. Somatic mutation analysis was performed with Strelka and Mutect version 1.1.7 (Cibulskis et al., 2013), excluding samples lacking a PASS designation. DNA structural variants were identified using Manta version 1.1.1 (Chen et al., 2016), requiring calls to bear a PASS or MGE10kb designation, tumor split read + tumor paired read ≥ 10, matched normal split reads = 0, and matched normal paired reads = 0. DNA copy number variants were identified using Canvas version 1.28.0-O01073 (Roller et al., 2016) and CopyCat (https://github.com/chrisamiller/copyCat) using the runPairedSampleAnalysis method with default parameters and performing GC correction. Copy number ratios were segmented by Circular Binary Segmentation implemented in the DNAcopy package (Olshen et al., 2004). RNA-seq analysis was performed using the Illumina RNAseq alignment app v1.1.0, aligning RNA FASTQ files to hg38-decoy using STAR version 2.5.0b (Dobin et al., 2013). Manual review of DNA and RNA data was performed using the Integrated Genomics Viewer (Robinson et al., 2011) (https://software.broadinstitute.org/software/igv) and the Illumina Variant Interpreter (https://variantinterpreter.informatics.illumina.com).

Mutation signature analysis

To perform per-sample mutation counting, all somatic mutations that were not excluded by quality filtering steps were counted. For mutation signature analysis this list was filtered using snpSift version 4.3g, including all alterations designated with the call “SNP” (Cingolani et al., 2012a). Evaluation of COSMIC mutation signatures was performed using the deconstructSigs package (Rosenthal et al., 2016), using the BSgenome.Hsapiens.UCSC.hg38 reference, the signatures.cosmic comparison set, a signature.cutoff value of 0.06, and a tri.counts.method parameter of “default”. De novo mutation signatures were derived using non-negative matrix factorization implemented in the SomaticSignatures R package (Gehring et al., 2015).

Evaluation of deletions with flanking microhomology

Deletions bearing microhomology were identified by a script counting deletions with two or more nucleotides of identical sequence between either 1) the 5’ end of the deleted region (determined from the HG38 genome reference) and the 3’ end following the deleted region or 2) the 3’ end of the deleted region and the 5’ end immediately preceding the deletion.

Evaluation of chromothripsis and chromoplexy

Chromothripsis was evaluated by counting the number of insertions, deletions, and copy number alterations within a moving 20 Mb. window positioned at 10 Kb. intervals along the entire genome, excluding telomeres and centromeres. Windows bearing at least 15 inversion rearrangements, 15 alternating copy number switches, and 10 deletions were called positive for chromothripsis. Chromoplexy was evaluated by applying the ChainFinder application version 1.0.1 (Baca et al., 2013) obtained from http://archive.broadinstitute.org/cancer/cga/chainfinder, using a deletion threshold of -0.278 and a significance threshold of 0.05. The presence of chromoplexy was defined by the presence of a chromoplexy chain connecting at least three chromosomes.

Noncoding mutation analysis

Recurrent promoter and untranslated region (UTR) point mutations were nominated by identifying mutations with variant allele frequency of at least 10% that were present in gene untranslated regions, enhancers, or promoters. UTR and promoter annotation were performed using ANNOVAR v2018Apr16 (Wang et al., 2010). Promoter regions were defined as 1 Kb upstream of the transcription start site. Enhancer regions were nominated by intersecting regions predicted by GeneHancer (Fishilevich et al., 2017) with regions enriched for H3K27ac histone modification identified by CHIPseq in (Kron et al., 2017). A peak in any of 19 samples in that data set was considered sufficient for inclusion in this analysis. Analysis of recurrent mutations in noncoding regions was restricted to regulatory regions predicted to affect any of the 574 genes listed in Tier 1 of the COSMIC Cancer Gene Census v85, obtained from https://cancer.sanger.ac.uk/cosmic/census (Futreal et al., 2004).

Data visualization and reporting

Circos plots were generated using the RCircos R package (Zhang et al., 2013).

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical Analysis

All statistical analysis was performed using R (v3.3.3) (R Core Team, 2018). Between-group comparisons of continuous variables were performed with the Wilcoxon rank sum test. Contingency table tests were performed with Fisher’s exact test. Correlation was assessed with Spearman’s correlation. All tests were two-sided.

DATA AND SOFTWARE AVAILABILITY

Sequencing Data

The raw sequencing data have been deposited in dbGAP under ID code (pending) (UCSF IRB approval obtained; NCI dbGAP ID is pending).

Supplementary Material

Supplemental Figure 1

Figure S1: DNA sequencing depth and quality metrics in normal and tumor samples, related to STAR methods. (A) Box and whisker plots showing distribution of mean coverage in normal, tumor. (B) Percentage of bases at sequencing quality ≥ Q30. (C) Total number of aligned reads. (D) Median insert length.

Supplemental Table 4

Table S4: Contingency table enumerating AR enhancer peak amplification and patient resistance to Abiraterone and/or Enzalutamide, related to Figure 2

Supplemental Table 5

Table S5: Mutations present in coding or regulatory regions associated with tumor suppressors and driver genes, related to Figures 1 and 5

Supplemental Figure 2

Figure S2: Genome-wide assessment of copy number and ploidy, related to Figure 1, table S1, Figure S1. (A) Mean copy number. Baseline copy number for chromosomes X and Y is 1 copy. (B) Mean genome-wide ploidy estimates. Estimated ploidy values for each sample are listed in Table S1. (C) Density plot for the number of mutations per megabase in tumors assigned diploid or triploid status as in (B). (D) Density plot for the number of translocations per megabase in tumors assigned diploid or triploid status as in (B).

Supplemental Figure 3

Figure S3: Structural variants affecting oncogenes and PTEN, related to Figure 1. (A) Schematic illustration of the PTEN gene locus, DNA copy loss (red lines), intersecting inversion rearrangements (green lines), translocations (black X). Below, locations of pathogenic missense or nonsense mutations (light/dark green circles) are indicated. Each variation was identified in a separate sample. (B) Expression of fusions activating oncogenes, sorted in increasing order by gene expression level, with somatic alterations noted below each gene. Red: amplification; Yellow: missense mutation; pink: gene fusion. See also Figure 1E listing upstream fusion partners. RNA expression measurements were available for 99 of 101 samples. RNA values are expressed as log(1+(TPM × 10^6). (C) RNA sequence of an ACPP-AXL fusion observed independently in a patient seen at the Vancouver Prostate Center; RNA generated from fresh frozen tumor tissue (radical prostatectomy) with high risk primary prostate cancer and methodology described in (Wyatt et al., 2014).

Supplemental Figure 4

Figure S4: RP11–356O9.1 is exclusively expressed in prostate, related to Figure 1. mRNA expression data expressed in RPKM obtained from GTEx as viewed on the UCSC genome browser (genome.ucsc.edu).

Supplemental Figure 5

Figure S5: Tandem Duplication length has a bimodal distribution in CDK12-inactivated tumors, related to Figure 3. Density plots of tandem duplication length, contrasting the three CDK12-mutant samples (yellow) with the 98 CDK12-WT samples (teal).

Supplemental Figure 6

Figure S6: Somatic signatures identified in mCRPC, related to Figure 4. (A) Trinucleotide context of signatures de novo 1 through 8. Signature de novo 2 was most likely due to a technical artifact, and was not considered during analysis. (B) COSMIC vs. de novo comparison of mutation signature fit. Results were hierarchically clustered.

Supplemental Table 1

Table S1: Molecular and clinical properties of the samples, related to Figures 1 and 3

Supplemental Table 2

Table S2: Regions with significantly elevated structural variant frequency, related to Figure 1

Supplemental Table 3

Table S3: Structural variations affecting key tumor suppressors and driver genes, related to Figures 1 and 5

Highlights.

• Deep whole genome and transcriptome sequencing of 101 prostate cancer metastases

• Tandem duplication affects intergenic regulatory loci upstream of AR and MYC

• Inactivation of CDK12, TP53, and BRCA2 affect distinct classes of structural variants

• Androgen receptor is affected by mutation or structural variation in 85% of mCRPC