Androgen receptor deregulation drives bromodomain-mediated chromatin alterations in prostate cancer

Summary

Global changes in chromatin accessibility may drive cancer progression by reprogramming transcription factor (TF) binding. In addition, histone acetylation readers such as bromodomain containing protein 4 (BRD4) have been shown to associate with these TFs and contribute to aggressive cancers including prostate cancer (PC). Here we show that chromatin accessibility defines castration resistant prostate cancer (CRPC). We show that the deregulation of androgen receptor (AR) expression is a driver of chromatin relaxation and that AR/androgen-regulated bromodomain-containing proteins (BRDs) mediate this effect. We also report that BRDs are overexpressed in CRPCs and that ATAD2 and BRD2 have prognostic value. Finally, we developed gene stratification signature (BROMO-10) for bromodomain response and PC prognostication, to inform current and future trials with drugs targeting these processes. Our findings provide a compelling rational for combination therapy targeting bromodomains in selected patients in which BRD-mediated TF binding is enhanced or modified as cancer progresses.

eTOC

Urbanucci et al. report how upregulated androgen receptor (AR) and AR-targeted bromodomain proteins contribute to genome-wide chromatin relaxation and increased gene transcription in advanced prostate cancer. They show that ATAD2 is a strong tissue biomarker and propose a BROMO-10 gene-signature to stratify patients for combination therapies with bromodomain inhibitors.

Introduction

Prostate cancer (PC) is the most common male cancer in the USA and Europe(Center et al.). Androgen Receptor (AR) signaling is required for the development of the prostate gland and is maintained in PC including at the stage of progression to castration resistant disease (CRPC)(Zhang et al., 2013).

CRPC is characterized by copy number gain at the AR locus occurring in around 30% of advanced cases. Consequently AR is overexpressed in these tumors. However, AR deregulation is also a frequent feature (>90%) of advanced castrate resistant cases (Waltering et al., 2012) which persist after resistance to antiandrogens such as enzalutamide and abiraterone (Buttigliero et al., 2015). We have previously shown that AR deregulation is associated with local chromatin landscape changes, which are able to reinforce the binding of AR to chromatin even in low androgen environments (Urbanucci et al., 2011; Urbanucci et al., 2012). This mimics the conditions occurring in CRPC (Xu et al., 2006). Phenotypically, AR deregulation results in increased growth rates even under conditions of androgen deprivation (Waltering et al., 2009). Moreover, genome wide AR recruitment to chromatin is detectable in such conditions (Andreu-Vieyra et al., 2011), suggesting that the chromatin is open and in some way primed by pre-docked AR even before the cells are treated with hormones. Androgen treatment then enhances AR recruitment (Urbanucci et al., 2011). This suggests that nucleosome positioning is predetermined in PC cells, a hypothesis that has been confirmed by a recent study (Chen et al., 2015).

Clinical epigenetics, defined as functionally relevant changes to the genome that do not involve a change in the nucleotide sequence and impact on disease phenotypes, is becoming extremely important for cancer detection and treatment. One indirect assessment of epigenetic alteration is the accessibility of DNA, determined by DNAse hypersensitivity analysis or by formaldehyde-assisted isolation of regulatory elements (FAIRE) (Song et al., 2011).

Proof of a de-regulated epigenome in CRPC includes altered patterns of DNA methylation, histone modifications (Perry et al., 2010) and increased and altered AR binding to chromatin (Pomerantz et al., 2015; Sharma et al., 2013).

In addition, epigenetic readers such as bromodomain containing protein 4 (BRD4) have been shown to associate with transcription factors (TFs) such as AR (Asangani et al., 2014; Nagarajan et al., 2014; Shi and Vakoc, 2014) and contribute to aggressive cancers of many types (Delmore et al., 2011; Shi and Vakoc, 2014) including PC (Asangani et al., 2014; Wyce et al., 2013). Nevertheless, the underlying mechanisms of action of bromodomain inhibitors have not yet been completely elucidated (Shi and Vakoc, 2014). Therefore we set out to investigate the underlying global changes in chromatin accessibility as a driver of cancer progression (Lever and Sheer, 2010; Timp and Feinberg, 2013).

Here we report that DNA accessibility alone is able to discriminate advanced prostate tumors from earlier disease states and benign tissue. The chromatin of these tumors is more accessible due to indirect mechanisms in which the AR plays a role. We show that bromodomain-containing proteins (BRDs) such as BRD4, and androgen-regulated BRD2 and ATAD2, are the mediators of such increased accessibility, and are prognostic tissue markers overexpressed in CRPC. Finally we provide a 10 gene signature, BROMO-10, that can be used to stratify patients with poorer outcome, and guide PC patient selection for combinatorial trials of bromodomain and extra-terminal (BET) - targeted therapies with other agents.

Results

Deregulation of the androgen receptor enhances bromodomain-mediated chromatin opening in advanced tumors

To understand whether progression to CRPC is associated with global changes in chromatin accessibility, we assessed chromatin opening and DNA regions with regulatory activity in 3 benign prostate hyperplasia (BPH), 3 primary untreated PC, and 3 locally recurrent CRPC samples using FAIRE-seq (Giresi and Lieb, 2009).

One-third of the genome consisted of open regions of chromatin in two CRPC samples whereas only about one sixth of the genome comprised open chromatin in the other samples (Figure 1A and Table S1A). We validated open chromatin regions via FAIRE-qPCR at PSA and TMPRSS2 gene loci, showing enhanced chromatin opening in CRPC (Figure 1B).

Figure 1. Deregulation of androgen receptor favors bromodomain-mediated chromatin opening in castration resistant prostate cancer. (See also Figures S1 and Tables S1A–S1D).

(A) Number of nucleotides located within open chromatin peaks identified via formaldehyde–assisted isolation of regulatory elements followed by sequencing (FAIRE-seq) in 3 benign prostate hyperplasia (BPH), 3 primary prostate cancer (PC) and 3 locally recurrent castration resistant prostate cancer (CRPC) tissue specimens. (B) FAIRE-qPCR validation of local chromatin opening at the PSA and TMPRSS2 loci in the clinical samples. An H3K27me3 marked region was used as negative control (closed region). (C) Overlap of open chromatin regions commonly found in BPH, PC, or CRPC samples according to the FAIRE-seq analysis. (D) Principal component analysis of 3 benign and 6 cancer (3 primary and 3 CRPC) prostate tissue samples according to chromatin shape. (E) Association between local chromatin opening, up to 1 kb upstream the genes transcription start site, and gene expression in matched tissue samples, according to RNA-seq. Analysis of Kolmorov-Smirnov statistical test values describing the deviation of correlative events between matched gene expression and local chromatin accessibility (F) or local DNA methylation (G) from a random set of associative events of the same type, in all 9 clinical samples. Local chromatin accessibility and DNA methylation were measured at the indicated intervals and correlated in a gene wise manner to respective gene expression in the same tissue (See Figure S4 for details). (H) Number of open chromatin regions found by FAIRE-seq analysis of LNCaP and VCaP cells treated with 1nM R1881 or vehicle for 4 hours. (I) FAIRE-qPCR validation of local chromatin opening at the PSA locus in LNCaP and VCaP cells treated for 4 hours with 1 nM R1881 or vehicle. Chromatin opening potency assessed by FAIRE-seq reads distribution around all FAIRE-seq sites (J) or around matched androgen receptor binding sites (ARBSs) (K) in LNCaP and VCaP cells treated as above. (L–N) FAIRE-qPCR analysis of local chromatin opening at the PSA and TMPRSS2 loci in a LNCaP-based AR overexpression model (* p<0.05 according to t-test). FAIRE-qPCR in LNCaP-pcDNA3.1 and -ARhi cells following four days of hormone starvation (l), or without starvation (full serum) (M), two of which with concomitant treatment with subtoxic concentration of 125 nM JQ1 or vehicle (DMSO) (* p<0.05 according to t-test). (N) FAIRE-qPCR upon transfection with siRNA control or siRNA against ATAD2 (siATAD2), BRD2 (siBRD2), or BRD4 (siBRD4) (2 biological repeats with three technical replicates each).

We built disease-stage specific high-confidence consensus open chromatin maps. On average, only 15% of the FAIRE-seq sites overlapped between the samples (Table S1A) highlighting epigenetic heterogeneity. We found that the majority of open sites were unique to CRPC samples (Figure 1C) and were larger than in PC or BPH (Figure S1A). To classify disease stage we used principal component analysis of the FAIRE-seq data and found that in terms of chromatin state, CRPC chromatin appeared more diverse when compared with BPH or PC (Figure 1D), indicating that extensive chromatin remodeling is a late event in PC progression.

RNA sequencing (RNA-seq) of the same clinical samples showed a positive correlation between upregulated genes and an open chromatin state (Figure 1E) up to 250kb upstream the transcriptional start sites (TSS) (Figure 1F). In agreement with published studies (Cedar and Bergman, 2012), DNA methylation profiles obtained for the same clinical samples showed a negative correlation with gene expression within 1kb and 5kb around the genes’ TSS (Figure 1G). Importantly, such correlations were independent of sample type, indicating conserved mechanisms across all disease stages.

Next, to investigate the effect of the AR deregulation on chromatin opening, we performed FAIRE-seq in CRPC cells such as LNCaP and VCaP cultured in presence and absence of androgens. VCaP cells overexpress the AR compared to LNCaP due to AR gene locus amplification (Urbanucci et al., 2013). More FAIRE-seq sites were found in VCaP than in LNCaP cells (Figure 1H and Table S1A) and VCaP cells also showed a greater increase in the number of FAIRE-seq sites in presence of androgens (Figure 1H).

One third of the FAIRE-seq sites were present in both cell lines (Figure 2A). Androgen treatment reprogrammed chromatin accessibility affecting 20 to 50 % of sites (Figure 2B).

Figure 2. AR overexpression in castration resistant prostate cancer cell models is associated with increased open chromatin.

Overlap of open chromatin regions found by FAIRE-seq analysis in LNCaP and VCaP cells treated with 1nM R1881 for 4 h or vehicle (etho=ethanol) (A&B). (c) FAIRE-qPCR validation of local chromatin opening at the TMPRSS2 locus (Urbanucci et al., 2011)) in two cell-lines with and without hormomne or vehicle. Chromatin opening around (D) androgen receptor binding sites (ARBS) in CRPCV (Sharma et al., 2013) (E) ERG binding sites (ERGBS) (Yu et al., 2010b) and (F) ARBS conserved between CRPC tissue and cell-lines (Sharma et al., 2013) (G) FAIRE-qPCR validation of chromatin opening at PSA and TMPRSS2 loci ((Urbanucci et al., 2011) in a LNCaP-based AR overexpression model .(Waltering et al., 2009)) treated with hormone or vehicle.

80% of common FAIRE-seq sites were conserved between cell lines and tissue samples (Table S1B). Chromatin opening measured via FAIRE-qPCR at PSA (Figure 1I) and TMPRSS2 (Figure 2C) loci, or measured in-silico genome-wide (Figure 1J and Figure 2D–F) was greater in VCaP cells than in LNCaP cells. An average of 21% of clinically relevant AR binding sites (ARBSs) in CRPC tissue (Sharma et al., 2013) and 45% ERG binding sites (ERGBSs) in VCaP cells (from (Yu et al., 2010a)) overlapped with FAIRE sites (Table S1B).

Androgens enhanced chromatin opening at FAIRE sites overlapping with CRPCs’ ARBSs and ERGBSs only in LNCaP cells (Figure 2, D and E) while had no effect in VCaP cells. In VCaP cells androgens enhanced chromatin opening at matched cell lines ARBSs (Figure 1K) and at ARBSs present in cell lines and tumors (Figure 2F) but not at ERGBSs (Figure 2E).

To validate the role of AR in chromatin opening, we used an LNCaP-based model expressing endogenous and increased AR levels (Waltering et al., 2009) (Figure 2G) and confirmed enhanced chromatin opening in AR overexpressing cells at the PSA and TMPRSS2 loci. These data suggest that both AR deregulation and androgens affect chromatin opening in CRPC cells at ARBSs, but not at ERGBSs.

Distribution of FAIRE-seq sites in clinical samples and cell lines showed increased opening at intronic/intergenic regions (Figure S1B–I), which is in agreement with previous findings showing increased AR chromatin binding at these regions (Sharma et al., 2013; Urbanucci et al., 2012). Given that chromatin remodeling is not exclusively associated with ARBSs, we tested whether enhanced chromatin opening was associated with presence of different motifs in CRPC by performing motif analysis on FAIRE site maps of tissue samples and cell lines (Table S1C). Consistent with its role in maintaining chromatin compaction (Tark-Dame et al., 2014), CTCF-like motifs were among the top enriched motifs in both clinical specimens and cell lines, followed by ETS-like motifs. CTCF and ETS motifs were enriched in all clinical specimens, including common sites, while c-MYC motifs were exclusively present in open regions found in CRPC samples. NFY and SP1 motifs were highly enriched in both treatment conditions in both cell lines. Although they were not enriched at FAIRE sites in tumors, NFY and SP1 have been shown to be involved in chromatin regulation and to have a potential role in cancer progression (Dolfini and Mantovani, 2013; Tewari et al., 2012). Interestingly, FOX-like motifs were significantly enriched only in the LNCaP FAIRE sites (Table S1C). This suggests that only a subset of FAIRE sites, those that may overlap with ARBSs, are regulated by FOXA1, and other chromatin remodelers may play a role. Therefore, we sought to understand whether enhanced chromatin opening in the context of AR deregulation favors chromatin binding of different proteins such as NFYA, the regulatory subunit of the NFY complex (Dolfini and Mantovani, 2013), and c-Myc.

We retrieved publicly available consensus binding data from ENCODE. Overlap of FAIRE-seq data with ENCODE data on chromatin binding of CTCF, MYC, and NFYA showed that, on average, 44% of the ENCODE NFYA sites, 18% of MYC sites and only 12% of CTCF sites lay within the open chromatin sites (Table S1D). However, no significant increase or decrease of overlapping sites was observed in CRPCs, which may be due to cell specificity of the TF binding or due to tissue heterogeneity (data not shown).

Bromodomain-containing proteins (BRDs) are druggable chromatin readers that recognize acetylated histones (Filippakopoulos et al., 2012) and modulate transcription in cancer associated genes (Filippakopoulos and Knapp, 2014; Lovén et al., 2013). Since we showed that histone acetylation is increased in AR overexpressing cells (Urbanucci et al., 2011), we hypothesized that BRDs would mediate chromatin opening. To test our hypothesis, we used the pan-BET bromodomain inhibitor JQ1 to assess the impact on chromatin opening potency at selected loci.

In hormone-starved cells, JQ1 reverted opening in AR overexpressing cells but had little or no effect in parental cells (Figure 1L). In the presence of androgens, JQ1 had a stronger effect on parental than on AR-overexpressing cells (Figure 1M), highlighting the role of androgens in chromatin opening. In addition to BRD2/4 as the main targets of JQ1, we also evaluated ATAD2, since it has been reported that this BRD is a common coactivator of AR and MYC.

RNAi knockdown of the BET proteins BRD2 and BRD4 reduced PSA enhancer and promoter opening in AR overexpressing cells, but in parental cells ATAD2 knockdown did not (Figure 1N). Interestingly, at the TMPRSS2 enhancer only BRD4 knockdown was effective in reducing chromatin opening, suggesting that compensatory events occur if a specific BRD is targeted. Based on these data we hypothesized that the expression of some BRDs might be AR dependent. To test androgen regulation of BRDs genes we validated proximal ARBSs (Supplemental Data File 1) identified in publicly available datasets. Chromatin immunoprecipitation at ATAD2 and BRD2 proximal ARBSs showed stronger AR binding than BRD4 proximal ARBS. The strongest ARBS to BRD4 was located 200kb from the TSS (Figure 3A). ATAD2 and BRD2 transcripts were induced by androgen (the latter only modestly), while BRD4 transcript was not (Figure 3B). AR overexpression sensitized ATAD2 transcription to lower concentrations of androgens (Figure 3C). In contrast to BRD4, ATAD2 and BRD2 protein levels were increased in AR-overexpressing cells and further increased by androgen stimulation (Figure 3D and Figure S2A). AR knockdown in LNCaP cells during a time-course of androgen treatment reduced ATAD2 transcript and protein levels within 24 hours (Figure S2, B and C). In contrast BRD2 protein levels were downregulated only after 48 hours treatment with AR-targeted siRNA (Figure S2C).

Figure 3. Androgens and androgen receptor regulate bromodomain containing proteins expression. (See Figure S2).

(A) Androgen receptor binding sites (ARBSs) close to BRD4, BRD2 and ATAD2 genes according to publicly available datasets (see Data Supplemental File 1) were validated by chromatin immunoprecipitation (ChIP)-qPCR analysis. PSA mid region served as ARBS negative control. (B) Indicated transcript levels measured by qRT-PCR after hormone treatment (C) ATAD2 gene expression measured by qRT-PCR in LNCaP-pcDNA3.1, -ARmo and -ARhi treated with hormone. The mean and S.E.M. of ATAD2 against TBP values normalized measure with no treatment (0M). (D) Western-blot analysis of indicated proteins levels in cells treated with hormone.

Bromodomain-containing proteins are tissue biomarkers overexpressed in castration resistant prostate cancers

To establish the predictive clinical value of BRD2/4 and ATAD2, we performed qRT-PCR (Figure 4, A–D) and immunohistochemical (IHC) analyses (Figure 4, E–H) using benign prostate tissue and prostate cancer specimens, and found that all transcripts were overexpressed in cancer compared to BPH. The long form of BRD4 (BRD4-L) (p>0.05) and BRD2 (p<0.0001) nuclear staining were increased in CRPC (Figure 4, I and J). BRD4 (Figure S3A) and BRD2 (Figure 4K) protein levels determined by IHC were not prognostic for biochemical recurrence, although BRD2 staining separated patients with poor prognosis, and was significantly (p=0.0154) associated with mortality (Figure 4L). Strong nuclear (Figure 4M) and cytoplasmic (Figure S3B) staining for ATAD2 was significantly increased in CRPC cases (p<0.0001). Moreover, strong nuclear (Figure 4N) but not cytoplasmic (Figure S3C) staining was associated with poor outcome (Figure S3D). We confirmed the significant (p<0.001) prognostic relevance of ATAD2 as a tissue biomarker for biochemical recurrence in an independent cohort of 12427 patients’ samples (Figure 4O and Figure S3). Positive staining for ATAD2 was associated with Ki67 staining, tumor stage, and AR protein expression (Figure S3).

Figure 4. Bromodomain containing proteins are tissue biomarkers overexpressed in castration resistant prostate cancer. (See Figure S3).

Expression of BRD4-long (BRD4-L) (A), BRD4-short (BRD4-S) transcript form (B), BRD2 (C), and ATAD2 (D) gene transcripts relative to TBP levels in BPH (n=15), primary untreated PC (n=27) and CRPC (n=15) specimens according to qRT-PCR. Kruskall-Wallis with Dunn post-test results are shown (***<0.0001** < 0.001; * 0.01 to 0.05; ns. not significant). CRPC immunohistochemical (IHC) strong stainings (score=3) for BRD4 long isoform (E) and BRD2 (F). CRPC IHC stainings showing examples of low (0%) (G) and high (<5%) (H) staining of nuclear ATAD2. Images are 8× magnification. Proportions of tumors according to BRD4 long isoform (I) and BRD2 (J) staining intensity in PC (n=159 for BRD4 and n=90 for BRD2) and CRPC (n=128 for BRD4 long isoform and n=34 for BRD2). Kaplan–Meier analysis of biochemical progression-free survival in 90 prostatectomy-treated patients according to BRD2 stainings (P= ns calculated with Mantel–Cox test) (K). Kaplan–Meier analysis showing shorter time to death in 37 men that died of PC out of the 90 patients for which material was stained for BRD2 (P=0.015 calculated with Mantel–Cox test) (L). (M) Proportions of tumors according to percentage of ATAD2 positive nuclei in PC (n=258) and CRPC (n=121) specimens (P<0.0001 according to X²test). (N) Kaplan–Meier analysis of biochemical progression-free survival in prostatectomy-treated patients according to the percentage of ATAD2 positive nuclei. 6 patients with high frequency of ATAD2-positive nuclei had very short progression-free time (P= 0.0354 calculated with Mantel–Cox test). (O) Kaplan–Meier analysis of biochemical progression-free survival in a validation cohort of 8541 prostatectomy-treated patients according to ATAD2 staining (p<0.001). 26% (n=2216) of the stainings were positive for ATAD2; of which strong ATAD2 staining accounted for 81% (n=1789).

Androgen receptor overexpressing cells are more sensitive to bromodomain inhibitors

BRD inhibitors have been reported to reduce the viability of PC cells (Asangani et al., 2014). To determine whether these effects are dependent on AR expression levels, we performed knockdowns of BRDs in AR-overexpression cell models in the presence of androgens (Figure S4, A and B). Silencing BRD4 decreased MYC levels in both LNCaP and VCaP cells, while upregulating slightly AR, and PSA only in VCaP cells (Figure S4B). Knockdown of BRD4 decreased viability in all the AR-positive cell lines tested (LNCaP-pcDNA3.1, LNCaP-ARhi, and VCaP), but knockdown of BRD2 or ATAD2 alone had no effect on viability (Figure 5A). ATAD2 inhibition via a small molecular probe was shown to have limited effect on viability of LNCaP cells (Bamborough et al., 2016). Silencing ATAD2 upregulated AR specifically in LNCaP and MYC in VCaP cells, while silencing BRD2 slightly upregulated MYC in both LNCaP and VCaP cells (Figure S4B). Therefore, our data suggest that compensatory mechanisms such as enhanced AR/MYC-signaling may take place and promote cell survival when single BRDs are targeted. In fact, co-targeting both ATAD2 and BRD2 in LNCaP cells via a combinatorial knockdown had no additive effect on decreasing cell viability compared to targeting BRD2 alone (Figure 5B), and highlighted that BRD4 is important but not the only contributor to cell viability.

Figure 5. Impact of Bromodomain inhibition on prostate cancer cell viability is enhanced by AR deregulation. (Refer also to Figure S4).

(A) Viability of LNCaP-pcDNA3.1, LNCaP-ARhi, and VCaP cells three days after transfection with siControl (siCont), siBRD2, siATAD2, or siBRD4, relative to control siRNA values. (B) Viability of parental LNCaP cells three days after transfection with siRNA as indicated. * indicates p<0.05 according to t-test versus the siCont column. N= 6 for each replicate, for each condition. Each experiment was repeated three times. (C) Relative viability of 22RV1, LNCaP-pcDNA3.1, LNCaP-ARhi, and VCaP cells cultured in full serum and treated with DMSO or JQ1(* p<0.05 according to t-test between the indicated conditions). (D) LNCaP-pcDNA3.1 and -ARhi cells were treated with 1nM DHT as well as JQ1 or vehicle (DMSO). Solo or combinatorial treatment of LNCaP-pcDNA3.1 and LNCaP-ARhi cells (E) or VCaP cells (F) with MDV3100 and JQ1. Viability compared to DMSO was assessed after 3 days treatment (* p<0.05 according to t-test versus JQ1 treatment alone). (G) Caspase activation assay upon treatment of VCaP cells with JQ1, MDV3100 or a combination, hydroxyl-urea was used as a positive control (H) Viability of MDV3100-resistant LNCaP cells treated with JQ1(* p<0.05 according to t-test versus DMSO).

To test whether AR deregulation enhances sensitivity to bromodomain inhibitors, we treated a panel of AR-positive PC cell-lines with JQ1 in presence of androgens (Figure 5C and Figure S4, C and D). AR levels in 22RV1 and LNCaP cells are similar (Erzurumlu and Ballar, 2017), while LNCaP-ARhi and VCaP cells overexpress AR compared to LNCaP (or LNCaP-pcDNA3.1) with VCaP showing the highest levels of AR (Urbanucci et al., 2012; Waltering et al., 2009). VCaP cells were indeed the most sensitive cells to JQ1 treatment. LNCaP-ARhi and parental LNCaP cells were equally responsive to JQ1 treatment (Figure S4D). However, when such cells were grown in androgen deprivation conditions (castrate conditions), AR overexpressing cells were more sensitive to JQ1 (Figure 5D). Also, combined treatment with enzalutamide and JQ1 was more effective in AR-overexpressing cells (Figure 5E and F) suggesting that AR deregulation is implicated in response to BET inhibition.

Combined treatment with JQ1 and enzalutamide triggered apoptosis in VCaP cells in full media (Figure 5G), but not in LNCaP cells, even when deprived of androgens (Figure S4E and F). This indicates that AR activity/level defines whether this drug combination has a cytostatic (low-level AR activation) or cytotoxic (high-level AR activation) effect.

When cells are treated with anti-androgens, resistance can emerge in which AR activity is maintained (Buttigliero et al., 2015). Using an enzalutamide-resistant LNCaP model (Kregel et al., 2013) we found that the inhibitory effect of JQ1 was retained (Figure 5H).

Gene expression analysis of bromodomain inhibitor–treated cells and six independent prostate cancer cohorts reveal a 10-gene signature for patient stratification

To identify patients that could potentially benefit from BET-targeted therapies, we sought to identify a gene-signature able to stratify CRPC responders. We performed a time-course treatment with JQ1 in LNCaP (Figure S5A) and VCaP (Figure S5B) cells to identify affected genes. Upregulated genes (Table S2A) showed significant overrepresentation of histone genes (Figure S5C and D) and overrepresentation of GO terms for chromatin compaction (p<10⁻⁶) (Table S3), corroborating the tendency for chromatin closure upon JQ1 treatment, while downregulated transcripts (Table S2B) included AR targets found in tissues of CRPC patients (Sharma et al., 2013) (Figure 6A). We validated decreased protein levels of the AR targets PSA and CAMKK2 upon JQ1 treatment (Figure S2A). CRPC-associated genes such as UBE2C, HOXB13, AURKA, and CAMKK2 (Supplemental Data File 1) that were downregulated by JQ1 treatment (Figure S5, E and F) and affected by BRD4 knockdown (Figure S5G), showed also bromodomain-dependent local chromatin opening (Figure S5H). AR deregulation affected chromatin opening especially at ARBSs, when present (Figure S5, I and J).

Figure 6. Bromodomain inhibition targets clinically relevant transcriptional program useful for selecting patients responsive to bromodomain-targeted therapies. (Refer also to Figure S5 and Tables S2–S5).

(A) GSEA of AR target gene signature (150 core genes identified in CRPC tissue) (Sharma et al., 2013) in expression analysis of VCaP cells treated with JQ1. (B) Heat map of genes displaying differential chromatin opening (varying distances upstream of TSSs) in CRPC vs primary PC. (C) Overlaps between overexpressed genes in CRPC in two clinical microarray datasets (Tables S4A and S4B), the genes associated with open chromatin sites in CRPC (Table S5) and the consensus genes downregulated by JQ1 treatment of two cell-lines (Table S2B) (D) 15 genes associated with open chromatin, overexpressed in CRPC and downregulated by JQ1 treatment. Underlined genes were also significant as assessed in Fred Hutchinson Cancer Research Centre (FHCRC) cohort comprising primary prostate cancers (PrCa; (Roudier et al., 2016)) and castration resistant PrCa (Kumar et al., 2016) (E) with indicated fold-change values and p values according to a two-sample t test (F)

Finally, we used published clinical expression array data (Taylor et al., 2010), and RNA-seq of clinical specimens (Ylipaa et al., 2015) to identify a clinical gene-signature of overexpressed genes in CRPC (Table S4A and Table S4B). We compared lists of CRPC overexpressed genes with those that displayed increased proximal chromatin opening in CRPC (Figure 6B and Table S5) and genes downregulated by JQ1 treatment in cell-lines, and obtained a set of 15 genes (Figure 6, C and D).

A generalized linear model with elastic net regularization (Erho et al., 2013) was used on the Mayo Clinic I (MCI) PC cohort (Erho et al., 2013) as the discovery dataset to assess the association of the 15 genes to biochemical recurrence, PC specific mortality, and metastatic recurrence. 10 of the 15 genes contributed significantly to the model.

We called the resulting ten-gene signature “BROMO-10”. To evaluate the prognostic significance of BROMO-10, first we used RNA-seq data from an independent cohort of PC from the Fred Hutchinson Cancer Research Center (Kumar et al., 2016; Roudier et al., 2016) (Figure 6E and F). Genes comprising the signature were deregulated in this cohort with 7 out of 10 genes being differentially expressed when comparing CRPCs to primary PCs.

Next, we assessed the independent prognostic value of each gene according to various endpoints in two additional validation cohorts (Karnes et al., 2013; Ross et al., 2015) (Table S6).

According to univariable and multivariable analysis, BROMO-10 contributed independent prognostic information over the clinicopathological variables. In the JHMI-RP cohort of high-risk men treated with radical prostatectomy without adjuvant or salvage therapy prior to metastatic onset, BROMO-10 discriminated for the biochemical recurrence endpoint (Figure 7A and Figure S6A and B), and the PC-specific mortality endpoint in the MCII cohort of high-risk men (Figure 7B and Figure S6C and D).

Figure 7. Prognostic and predictive value of the 10 genes signature BROMO-10. (See Figure S6 and Table S6).

Kaplan Meier curves indicating the prognostic separation achieved by high and low BROMO-10 scores versus (A) biochemical recurrence in the JHMI-RP validation cohort(Ross et al., 2015) (p value 0.008) (B) PCSM free survival in the Mayo Clinic validation cohort (Karnes et al., 2013) (p value0.0089) and (C) CRPC-free survival of patients post-ADT treatment of metastatic patients in the JHMI cohort as expression quartiles (n=55, p value 0.01). (D) Heatmaps of BROMO-10 score and individual gene expression in cell-lines and also in PDX models (E). Two-sample t test evaluation of the significance of growth inhibition or reduction in tumour volume with IC50 dose administration of BET bromodomain inhibitors JQ1 (F), ZEN-3694 (G), I-BET762 (H) to cell-lines or of ZEN-3694 to PDX xenografts (I).

We further assessed the prognostic value of BROMO-10 in predicting the onset of CRPC. Fifty-five patients that developed metastasis after radical prostatectomy without any adjuvant or salvage therapy ("natural history cohort") treated at Johns Hopkins were evaluated using Kaplan Meier analysis considering time to CRPC from metastatic onset. Upon metastatic onset all patients received androgen deprivation therapy (ADT). Higher BROMO-10 scores were associated with an increased rate of CRPC after ADT (Figure 7C). The upper quartile of BROMO-10 signature scores had a median time to CRPC of 12 months compared to the lower quartile of 84 months (p = 0.01).

Finally, to assess whether BROMO-10 is able to predict responsiveness of PC to bromodomain inhibitors, we used RNA-seq profiles of panels of PC cell lines and expression array profiles of patient-derived xenografts (PDXs) (Nguyen et al., 2017) to generate BROMO-10 scores (Figure 7 D and E). We then correlated these scores with publicly available IC50 data for JQ1 (Asangani et al., 2016) (Figure 7F), ZEN-3694 (Attwell et al., 2016) (Figure 7G), I-BET762 (Figure 7H), and responsiveness to I-BET762 treatment measured as a significant reduction in tumor volume (Wyce et al., 2013) (Figure 7I). A high BROMO-10 score was significantly (p<0.001) associated with responsiveness of PC cell lines to JQ1 and I-BET762, and reduction of tumor volume upon treatment of PC PDX models with I-BET762 (p<0.01). However, the comparison between groups was not significant for responsivness to ZEN-364 because PC3 cells, notably an AR negative PC cell line, had low BROMO-10 score and was nevertheless sensitive to ZEN-364. Taken together these data confirm the ability of BROMO-10 to predict responsiveness to BET inhibitors.

Discussion

Here we show that chromatin accessibility increases during PC cancer progression due to mechanisms that involve the AR overexpression and the activity of BRDs. Importantly, we found that the chromatin structure in CRPC is able to classify disease stage demonstrating that genome-wide chromatin structure is reprogrammed as disease progresses and it shows distinct features compared to primary tumors or benign tissue. Increased chromatin accessibility in PC was inferred in a recent study although the low number of peaks found in the healthy tissue dominated the results (Stelloo et al., 2015). For the present study we developed an advanced analysis pipeline to exclude possible confounding factors such as variations in ploidy from sample to sample (see supplemental information). This was essential for improved analyses because copy number variation has previously been shown to be strongly associated with poor prognosis in advance disease (Taylor et al., 2010).

Interestingly androgens were able to enhance chromatin opening especially at ARBSs. This suggests a positive feedback loop in which the AR is able to bind more tightly to the genome due to increased opening at ARBSs. These data are concordant with our previous results showing stronger AR binding to chromatin in AR-overexpressing cells (Massie et al., 2011; Urbanucci et al., 2012), and to other reports showing that chromatin accessibility is pre-docked prior AR binding at ARBSs (Andreu-Vieyra et al., 2011; He et al., 2010).

BRDs have gained extensive attention due to their tissue-specific capacity to modulate key transcriptional events during cancer progression (Fu et al., 2015) also in CRPC where they are therapeutically relevant (Asangani et al., 2014). We found that selected key BRDs such as ATAD2, BRD2, and BRD4 have a locus-specific effect on chromatin opening, suggesting that compensatory mechanisms may take place if BRDs are inhibited with single agents. For instance, we show that upregulation of the AR or MYC proteins occurs while inhibiting ATAD2 or BRD2, and possibly explain the limited effect of their inhibition on cell viability.

We also show that ATAD2 and BRD2 are androgen regulated. ATAD2 was previously reported to be androgen regulated (Zou et al., 2009). However, here we report that ATAD2 expression is enhanced in AR-overexpressing cells at low concentrations of androgens, and BRD4 long isoform, BRD2, and ATAD2 are all overexpressed in CRPC tissues. These results support the presence of an AR deregulation-mediated positive feedback loop that boosts the expression of BRDs in order to increase AR chromatin accessibility. Moreover, ATAD2 had strong prognostic value on a cohort of ten thousand patients. Our data also suggest that while ATAD2 is an optimal tissue biomarker in identifying PC tissues where active transcription due to heavy cell cycle turnover is ongoing, it may not be a good target for PC therapy, as also shown recently (Bamborough et al., 2016), if not targeted in combination with other agents such as antiandrogens.

The role of ATAD2 as a regulator of chromatin dynamics is well known in yeast (Cattaneo et al., 2014). A recent study showed that ATAD2 is highly expressed in replicating PC cells and ATAD2 expression correlates with expression of cell cycle and DNA replication genes that have overlapping function in meiosis and tumor progression (Koo et al., 2016). Moreover, in highly proliferating embryonic stem cells, ATAD2 was reported to sustain specific gene expression programs via regulating chromatin opening guided by histone acetylation (Morozumi et al., 2016), which is in agreement with our data. These findings corroborate our data and are supportive of ATAD2 being a possible contributor to increased transcription plasticity in CRPC.

Our study also suggests that there might be a subpopulation of tumors that are more dependent on BRDs activity than others. Therefore, we built a 10 gene signature, BROMO-10, which is able to discriminate patients with poorer outcome, which takes into account chromatin structure and additionally incorporates key PC-specific downstream targets of BRDs. Interestingly these targets include FEN1, which we have previously described to be important for PC progression and proposed as a tissue biomarker for biochemical recurrence (Urbanucci et al., 2012), EEF1A2 which has been proposed as a marker of prostate cell transformation (Scaggiante et al., 2012; Sun et al., 2014), KAT2A, which encodes a histone acetyl-transferase controlling the PI3/AKT pathway with therapeutic potential in leukemia (Sun et al., 2015), and HSPH1 which enhances MYC transcription and drives B-cell non-Hodgkin lymphoma (Zappasodi et al., 2015).

Mechanisms of resistance to BET inhibition have been reported. Therefore, it is extremely important to define patients that will respond to BET-inhibition therapies in combination with standard therapies to avoid resistance. We show that BROMO-10 is able to predict response to BET-inhibition therapies, but the use of this signature should be limited to tumors with intact AR signaling and further studies are needed to refine the signature for different compounds.

In conclusion, we propose AR deregulation-driven chromatin structure as a key determinant of tumor progression. We describe the effect of BET inhibition on chromatin accessibility as an additional mechanism by which it is able to repress cell growth in a cell-specific manner. Moreover, we propose BROMO-10 signature to be used to select patients more likely to benefit from BET-targeted therapies and avoid recurrence. The selection of PC patients into future trials evaluating the efficacy should be based on the assessment of AR status, key BRDs expression, such as ATAD2, and the gene signature that reflects the chromatin status of these tumors.

Experimental Procedure

Clinical samples

All work on clinical samples has been carried out in compliance with the Helsinki Declaration and with the approval of ethics boards at collaborating institutions as outlined below.The tissue microarray from the Department of Urology and the Martini Clinics at the University Medical Centre Hamburg-Eppendorf consisted of archived diagnostic leftover tissues. Manufacture and analysis was approved by the local ethics committee (Ethics commission Hamburg, WF-049/09 and PV3652). According to local laws (HmbKHG, §12,1), informed consent was not required for this study. Patient records/information was anonymized and de-identified prior to analysis.

3 BPH, 6 primary PC and 3 CRPCs were used for FAIRE-seq assays and FAIRE-qPCR assays. RNA-seq data from transcriptomes of 12 BPH, 28 untreated PCs, and 13 CRPCs, including the samples used for FAIRE-seq, were publicly available (Ylipaa et al., 2015). These samples and the tissue microarray described below were provided by Tampere University Hospital. The use of these samples for FAIRE-seq and of the tissue microarray was approved by the ethical committee of Tampere University Hospital and the National Authority for Medicolegal Affairs. Written informed consent was obtained from the subjects for sequencing the samples.

The Tampere patients’ cohort of tissue microarrays (TMAs) contained 258 formalin-fixed paraffin embedded prostatectomy and 121 CRPC specimens. Subset of the cohort was used to immunostain for ATAD2, BRD2, and BRD4. The Hamburg patients’ TMA cohort contained 9467 prostatectomy tissue specimens. Radical prostatectomy specimens were available from 12,427 patients. PSA values were measured following surgery and PSA recurrence was defined as the time point when postoperative PSA was increasing from at least 0.2ng/ml.

Formaldehyde-Assisted Isolation of Regulatory Elements (FAIRE)

Two replicates were processed for each cell line and condition for subsequent sequencing analysis. 3–5 replicates were processed for qPCR analysis. Four million cells were plated and hormone-deprived for 4 days. Cells were then treated with R1881 or DHT for 4 hours. To perform tissue FAIRE from clinical material 3 ml of PBS containing 2× protease inhibitor (Roche Inc., Mannheim, Germany) were added to 40×20 µm sections of freshly frozen tissue specimens. Downstream fixation and processing of both sample types is as described in the Supplemental Information.

FAIRE-seq analysis

Peak detection for FAIRE-seq was performed using MACS (Zhang et al., 2008) with default parameters using inputs of each of the FAIRE samples as controls, and with F-Seq (Boyle et al., 2008). Refer to Supplemental Information for more detail.

Evaluation of the 15-gene signature prognostic value

Microarray data from the Decipher GRID™ were extracted for three radical prostatectomy cohorts from previously described (Erho et al., 2013; Karnes et al., 2013; Ross et al., 2015) validation studies. A classifier to distinguish between metastatic vs. non-metastatic cancers was developed. The classifier was constructed from 15 genes (Figure 5C) using a generalized linear model with elastic net regularization as previously described (Erho et al., 2013). The model was generated using the MCI cohort (GSE46691) as training data. In the final model 10 of the 15 genes contributed to the model score with 6 positively associated and 4 negatively associated genes as determined by the regularized coefficients (see also Table S6). Scores were then generated for samples from the Mayo Clinic II (MCII) and JHMI-RP validation cohorts and performance in each cohort was assessed using survival analysis.

BROMO-10 scores in cell lines and patients derived xenografts were calculated using GSVA Bioconductor package (https://www.bioconductor.org/packages/release/bioc/html/GSVA.html).

Statistics

Statistical analyses were performed using GraphPad Prism, MATLAB, and Microsoft Excel. All statistical tests were two-tailed with testing level thresholds of α = 0.05.

Supplementary Material

2

Supplemental Data File 1. Locations of androgen receptor binding sites and FAIRE-seq sites proximal to genes of interest. Relates to Figure 3.

3

Supplemental_Info_Figures_S1_S6_and_Experimental_Procedures: Figures S1 to S6 and Supplemental Experimental Procedures.

4

Table S1: Overview of the numbers of properties of open chromatin sites identified in clinical samples and cell-lines (related to Figure 1). Open chromatin sites in (A) clinical specimens and cell-lines (B) overlapping with publicly available datasets (C) subjected to motif analysis at sites overlapping between clinical samples and cell-lines (D) overlapping with ENCODE datasets.

5

Table S2. Genelists for genes differentially expressed in response to JQ1 treatment (related to Figure 6). Upregulated genes (A) and downregulated genes (B) by JQ1 after 5, 12, or 24 hours treatment of LNCaP and VCaP cells with 1nM R1881 and 125nM JQ1 versus 1nM R1881 alone.

6

Table S3: Gene Ontology enrichment analysis (p<10-6) of affected genes upon 5, 12, and 24 hours treatment with JQ1 in LNCaP and VCaP cells, Related to Figure 6.

7

Table S4. Overexpressed genes in CRPC versus primary PC or BPH samples from two published datasets (related to Figure 6). (A) Taylor et al., dataset (log2FC>0.4 including both CRPC vs PC and CRPC vs BPH) and (B) Hylipaa et al. RNAseq dataset (log2FC>0.4 including both CRPC vs PC and CRPC vs BPH)

8

Table S5. List of genes that displayed increased proximal chromatin opening between CRPC and PC according to FAIREseq data, Related to Figure 6.

9

Table S6: 10 genes signature made of contributors with non-zero-coefficients according to a generalized linear model with elastic net regularization in the training (MCI) cohort and individual prognostic values (p-value according to Kaplan Meyer analysis) in the training and discovery datasets (JHMI-RP and MCII cohorts), Related to Figure 7.

Highlights.

• Androgen Receptor (AR) overexpression drives genome-wide chromatin relaxation

• Bromodomain (BRD) inhibitors reduce local chromatin opening at AR target genes

• AR overexpression upregulates BRD proteins ATAD2, BRD2 and BRD4

• The BROMO-10 gene signature stratifies patients for therapy with BRD inhibitors