Background
Prostate cancer remains the most commonly diagnosed cancer in men, with a lifetime risk of approximately 1 in 7 [
1,
2]. Both cancer-associated events and the normal physiology of prostate involve signaling through the androgen receptor (AR) [
3]. Indeed, clinical intervention based on androgen deprivation therapy (ADT), which reduces AR signaling, is a cornerstone of prostate cancer treatment. Resistance to ADT invariably develops and leads to the development of castrate resistant prostate cancer (CRPC) with associated morbidity and mortality. CRPC is characterized by changes in growth factor-, cell surface receptor-, and kinase-dependent signaling as well as gene expression that impact fundamental processes such as cell growth, motility, and DNA repair [
4,
5]. Understanding how these changes occur, and defining actionable targets within the affected pathways, could expand the options for treating CRPC.
DNA repair enzymes have emerged as actionable targets for cancer, including prostate cancer. The data from clinical trials has shown that inhibiting the DNA repair enzyme PARP-1 in ovarian, breast, and prostate cancers improves outcome in patients that have genetic alterations in other components of the DNA repair machinery [
6‐
8]. The success of PARP-1 inhibitors in this context suggests that new therapeutic opportunities might be revealed by understanding the interplay between genomic status and DNA repair pathways. There are also strong indications that the response to ionizing radiation (IR) can be influenced by androgen signaling. Thus, in pre-clinical models and in patients, ADT can confer radiosensitivity [
9‐
11]. Finally, it has been shown that treating prostate cancer xenografts with inhibitors to AR (bicalutimide) and PARP-1 (Olaparib) inhibits tumor growth [
12].
Pre-clinical models of prostate cancer, particularly a relatively small number of cell lines, are widely employed to study signal transduction and transcription, and to evaluate drug and IR sensitivities. A critical knowledge gap that may limit the interpretation and possibly the impact of data generated from these models is the genomic and transcriptomic state of the cells. To address this gap, we have performed whole genome sequencing (WGS) and RNA-seq on three prostate cancer lines. Combining the new data with publicly available data from human prostate cancers, we explored two important issues related to DNA damage signaling and repair. The first question was, do prostate cancer models harbor deleterious mutations in the DNA repair machinery. The second question was, does androgen signaling regulate expression of DNA repair machinery.
Our analysis revealed the presence of missense mutations in DNA repair genes in LNCaP, VCaP, PC3-AR, and RWPE-1 cells. Across these models, a total of 34 DNA damage response (DDR) genes were up-regulated, and 87 DDR genes were down-regulated in response to androgen. By co-expression network analysis, we found that expression of 25 DDR genes were altered in response to androgen treatment of cell lines. We also explored the interplay between DNA repair and AR-dependent transcription. Treating cells with the small molecular inhibitor, mirin, which inhibits the endonuclease MRE11, reduced AR-induced transcription. Our genomic and RNA-seq data should be useful for groups studying how the status of the DNA repair machinery influences properties such as drug sensitivity. The mirin effect on AR activity and cell growth suggests it might have utility as an inhibitor of prostate cancer cells.
Methods
Antibodies, reagents, and standard techniques
For immunoblotting experiments, lab-prepared AR hinge 3 (against AR residues 656 to 669: TQKLTVSHIEGYEC), Hsp70 (Stressgen), Hsp90β (GeneTex), lab-prepared FKBP51 (against FL protein), and Tubulin (Sigma Aldrich) were used. Secondary antibodies were IRDye-800 labeled antibodies (Rockland #610–732- 124, #610–132-121, and #611–732- 127), and Alexa Fluor-680 labeled secondary antibodies (Life Technologies #A21058 and #A10043). Standard immunoblotting procedures were conducted, images were detected on a fluorescent Odyssey imager, and analyzed using the provided LI-COR software.
For immunofluorescence, a lab prepared AR21 antibody (against AR residues 1 to 21: MEVQLGLGRVYPRPPSKTYRGC) was used while DAPI staining was used for nuclear detection. The secondary antibody was a Cy3-labeled anti-rabbit (Jackson Immunoresearch). Cells were grown on glass coverslips. Coverslips were prepared using standard immunofluorescence methods with a 15-min fixation (3.75% formaldehyde), washed with PBS, permeabilized for 5 min (0.2% triton X-100), and incubated with a 1-h block at room temperature (2% FBS and 2% BSA in PBS). Coverslips were then incubated in primary antibody (diluted in blocking buffer) overnight at 4 °C. Secondary antibodies (diluted in blocking buffer) were incubated for 1 h at room temperature. Images were obtained using a confocal microscope (Zeiss 800 LSM, Carl Zeiss) at 40×, 1.3 NA oil immersion objective and captured/processed using ZEN software (Carl Zeiss).
For immunoprecipitation, cells were lysed with cell lysis buffer (20 mM Tris-HCl pH 7.5, 50 mM NaCl, 0.5% Triton X-100, 5 mM EDTA, 2 mM DTT, and protease inhibitors) and clarified by centrifugation prior to incubation with Anti-FLAG M2 Affinity Gel (Sigma-Aldrich) at 4 °C. After a 4-h incubation, beads were washed with wash buffer (20 mM Tris-HCl pH 7.5, 50 mM NaCl, 0.1% Triton X-100, 0.1 mM EDTA, 2 mM DTT, and protease inhibitors). Standard SDS-PAGE loading buffer and procedures were used to separate proteins.
Cell culture
LNCaP, VCaP, and PC-3 cells were kindly provided by Dr. Michael Weber (University of Virginia) and were purchased originally from ATCC. VCaP cells were grown in DMEM supplemented with 10% FBS and 1% antibiotic/antimycotic. LNCaP and PC-3 cells were grown in RPMI supplemented with 5% FBS and 1% antibiotic/antimycotic. PC3-AR cells were made by stable lenti-viral infection of full length AR using a pWPI-GFP-FLAG-AR plasmid in which the GFP portion was swapped with the antibiotic selectable hygromycin resistance gene. RWPE-1 cells were obtained from Dr. Daniel Gioeli and cultured in Keratinocyte SFM supplemented with the provided EGF and BPE factors. All cells were incubated at 5% CO2 and 37 °C.
Cell growth/survival assays and cell cycle analysis
Cells were seeded onto a 96-well format for 1 day. Media was exchanged and supplemented with indicated concentrations of inhibitors for 72 h. Alamar blue dye (Promega, #G808A) was added (10% of total volume) for ~ 6 h and measured with a fluorescent plate reader according to manufacturer’s recommendations. Technical replicates of at least 4 measurements were averaged. The data was normalized by removing the background signal and rescaling the values so that the vehicle condition was 100%. Values were plotted using Prism software and IC50 values were calculated with non-linear regression based on the log transformed data.
Cell cycle measurements were determined using the FITC-conjugated BrdU Flow Kit (BD, #559619) where we stained and performed flow cytometry of asynchronous cells according to the manufacturer’s instructions and in reference to Benamar et al. (2016) [
13]. In brief, cells were pulsed with BrdU for 1 h. Cells were washed and harvested using trypsin. After a PBS wash, cells were fixed. Prior to staining with 7-AAD and anti-BrdU antibody, cells were permeabilized. Cells were resuspended in a 1-mL solution and more than 10,000 cells were measured using the Cytek modified BD FACSCalibur provided by the Flow Cytometry Core Facility at the University of Virginia. Data was analyzed using the FlowJo and ModFit software packages.
Gene expression analysis
Prostate cancer cells were plated in the corresponding phenol-free based media supplemented with charcoal-stripped serum (Gemini) for 48–72 h. Synthetic androgen, R1881 (Sigma Aldrich), was typically added for 12 h at a concentration of 2 nM.
For real time quantitative PCR (RT-qPCR) experiments, RNA was extracted using standard TRIzol (Thermo Fisher Scientific) methods. cDNA was prepared using BioRad iScript reagents and expression was detected using SensiMix Sybrgreen reagents, all according to manufacturer’s instructions. Technical replicates were averaged and normalized to the GUS housekeeping gene. Experiments are representatives of at least 3 experiments.
The following primers were used at a final concentration of 200 nM:
For RNA-sequencing experiments, the Qiagen RNeasy kit was used to extract RNA. Library preparation and sequencing was performed by Hudson Alpha. Briefly, RNA integrity and concentration were assessed by a fluorometric assay, indexed libraries were made using the standard polyA method, quality control was used to determine size and concentration, and samples were sequenced using Illumina HiSeq 2500 at a depth of 250 million × 50-bp paired-end reads. Reads were aligned to the hg38 genome (ENSEMBL GRCh38.89) using STAR (release v. 2.5) [
14]. Counts were generated using HTSeq (release v. 0.6) [
15]. DESeq2 R package was used to determine normalized counts [
16]. Genes with low counts were eliminated (≤ 10 in all conditions), and definitions of differential genes are described in the figure legends.
For weighted gene co-expression network analyses (WGCNA), we filtered the count matrix to remove genes with low read counts (where sum of reads in all samples < 1). We then applied variance stabilizing transformation to the remaining data resulting in homoskedastic counts normalized with respect to library size. Unsupervised clustering was performed with WGCNA [
17,
18]. Briefly, a network was constructed using biweight midcorrelation as the measure of similarity between genes with β equal to 5. Modules were identified by applying hierarchical clustering (average method) to distance calculated from signed topological overlap matrix and the tree was cut with cutreeDynamic using the following parameters: minimum module size equal to 30 and hybrid method. Next, the modules were merged if the distance between them was equal to less than 0.25, resulting in 15 modules. We then calculated the eigengene for those 15 modules and created a gene list representing each module by filtering the genes based on gene significance and intra-modular connectivity. Modules were subsequently described by overrepresented pathways using Enrichr. Gene Set Enrichment Analysis (GSEA) was performed on pre-ranked gene list that was generated by assigning a value to each gene that was equal to log of
p-value multiplied by the negative sign of the fold change (rank = − 1 * sign(FC) * log(
p-value)). Gene sets used for analysis with GSEA included the MSigDB hallmark gene sets [
19] as well as a curated DDR gene set of 450 genes.
Data sources
Androgen-dependent and androgen-independent microarray data was downloaded using Gene Expression Omnibus (GEO) data repository (GSE847) [
20]. We restricted our analysis to a curated list of DNA repair genes [
21]. To confidently assess the expression of DNA repair genes, we filtered the normalized data by re-scoring negative intensities and values with “absent” detection calls [
22]. Fold changes between hormone insensitive and hormone sensitive xenograft pairs were calculated, and fold changes ≥1.5 and ≤ 0.667 were defined as an alteration.
The publicly available Memorial Sloan-Kettering Cancer Center (MSKCC) prostate adenocarcinoma (PRAD) dataset was downloaded from cBioPortal [
23,
24]. Clinical and expression data for a total of 181 primary prostate tumors and 37 metastatic tumors were provided by Taylor et al. (2010) [
25]. We limited our analysis to the 450 expert-curated DNA damage/DNA repair gene set from Pearl et al. (2015) [
21] and tested the association of the normalized expression for various clinical parameters using the Kruskal-Wallis test (
p-value ≤1E-5). To generate the heatmap, we used “mRNA Expression Z-Scores vs Normals” data that was normalized and analyzed by cBioPortal. We plotted those genes found to be significantly associated with disease status using ComplexHeatmap.
In order to test a significance based on DNA damage response ontologies, we used Fisher’s exact test (fisher.test() in R v. 3). We generated unions of genes from each of the 125 ontology pathways described by Pearl et al. (2015) [
21]. We then classified the genes from each of our analyses (cell line, xenograft, patient metastases, or all groups) based on their presence or absence in the ontology group. Returned
p-values for each ontology group were plotted as a function of the -log10 value.
ChIP-sequencing analysis was derived from GSE28126. LNCaP and VCaP data was aligned to the hg38 reference genome using bowtie2 [
26]. Peaks were called using macs2 [
27]. To determine androgen-induced peaks, R1881 treated samples were analyzed as the “treatment” while untreated samples were analyzed as the “control”. Peaks were annotated to the closest gene using bedtools [
28].
Additional image processing information
Figures related to RT-qPCR and S-plots were generated using Microsoft Excel and graphed using GraphPad Prism (GraphPad Software, La Jolla, CA). GSEA (Broad Institute, Inc., release v. 3.0) was used to calculate gene set enrichments [
29,
30]. The data was replotted using R. ComplexHeatmap and Python (matplotlib.pyplot.imshow) were used to graph heatmap data. Lollipops (release v. 1.3.2) was used to plot protein mutations [
31]. All figures were assembled with Adobe Illustrator.
Detecting genomic variants
For LNCaP, VCaP, and PC3-AR cell lines, genomic DNA was prepared using the Qiagen DNeasy kit. Libraries were prepared and samples were sequenced by Hudson Alpha. Sequenced DNA data from RWPE-1 was kindly provided by Dr. Anindya Dutta (University of Virginia). Generation of sequenced RNA datasets were described above. To provide additional coverage for RNA transcripts, the sequenced reads from both the control and the androgen-treated samples were merged for each respective cell line. The RWPE-1 RNA-seq sample was derived by merging the following publicly available datasets: SRR1282953, SRR2919800, and SRR2919800.
DNA was aligned using BWA [
32,
33] to the hg19 reference sequence. Corresponding RNA-seq was aligned using STAR [
14] to the hg19 reference sequence. Aligned reads were subsequently filtered and processed using GATK Haplotype Caller [
34]. Variants were limited to the following 18 DNA repair genes:
PARP1,
PARP2,
ERCC3,
ATR,
ATM,
RAD50,
RAD51,
MRE11,
NBN,
CHEK1,
CHEK2,
MLH3,
PALB2,
FANCA,
BRCA1,
BRCA2,
HDAC2, and
PRKDC. Allelic frequency for each variant were compared to the 1000 Genomes Project [
35] and the NHLBI GO Exome Sequence Project [
36]. COSMIC identification numbers [
37,
38] and prior reports for each variant was verified through a literature search [
39]. Deleterious mutations were predicted in silico by Scale-invariant feature transform (SIFT) [
40], fitness consequence (fitCons) [
41], Combined Annotation-Dependent Depletion (CADD) [
42], and Polymorphism Phenotyping (PolyPhen) [
43].
Discussion
The discovery that PARP inhibitors have efficacy in prostate cancer patients that harbor mutations in DNA repair genes provides proof-of-concept that targeting the DNA repair machinery can be beneficial. Defining the genomic status and RNA expression of the DNA repair machinery in prostate cancer cell lines provides a knowledge base that is critical for the selection of an appropriate model, and interpretation of data generated from the model. This could include evaluating the effects of clinically-used drugs, screening for new compounds, and searching for potential synthetic interactions. In this study, we set out to understand the genetic changes in the DNA repair machinery in LNCaP, VCaP, PC3-AR, and RWPE-1 cell lines. We found a total of 24 small nucleotide variants with low allelic frequency across these prostate cancer cell lines. Mutations were detected in the following genes:
ATM,
ATR,
BRCA1,
BRCA2,
CHEK2,
PRKDC,
ERCC3,
FANCA,
HDAC2,
MLH3,
NBN,
PARP1, and
RAD50. Of these 24 mutations, two mutations in LNCaP cells are predicted to be deleterious CHEK2(E239*) and RAD50(L719 fs*15). The
RAD50 mutation (chr5:131931452; L719 fs*15) causes a frameshift and a large C-terminal truncation that results in loss of > 500 amino acids. The
CHEK2 mutation (chr22:29107974; E239*) introduces a stop codon that deletes the kinase domain. CHEK2 activation in response to DNA damage induces a cell cycle checkpoint [
71].
CHEK2 variants predispose individuals to breast and colon cancer [
72] and it has been shown to be a negative regulator of prostate cancer growth [
73]. RAD50 is a member of the MRN (MRE11-RAD50-NBS1) complex which functions as a scaffold for sensing DNA damage [
74]. Mouse knockouts of
RAD50 are embryonic lethal [
75], and, like
CHEK2,
RAD50 mutations are associated with cancer risk [
76,
77]. Mutations in
CHEK2 or
RAD50 could help sensitize LNCaP cells to DNA damage induced by chemotherapy drugs and IR. From sequence data, the genomic alterations in
CHEK2 and
RAD50 appear to affect single alleles in LNCaP cells, but
CHEK2 and
RAD50 are both haploinsufficient genes, the level of activity provided from a single WT allele may not provide enough activity for the cell.
Additionally, we detected missense mutations in two additional DNA repair genes that are predicted to impact protein function. The
ERCC3 mutation (chr2:128044450; R391W) introduces a tryptophan into the Helicase C-terminal domain, and the
PARP1 mutations (chr1:226566948; E547G and chr1:226555302; V762A) introduce amino acid changes into functional domains of the protein. One of the amino acid changes, V762A, has been studied biochemically and shown to reduce PARP-1 enzyme activity [
78,
79].
Our characterization of DNA repair genes in prostate cancer cells included generating RNA-seq data, determining the extent to which DNA repair genes are regulated by androgen in different cell lines, and using existing data sets to make comparisons with xenograft models and human prostate cancer. Androgen signaling through AR has been shown to induce a DNA repair signature (32 genes) that could explain the radioresistance of some prostate cancers [
50].
Using RNA-seq data and WGCNA, we found that 17 DNA repair genes were positively correlated with androgen treatment across three cell lines, and eight DNA repair genes were negatively correlated with androgen treatment. Surprisingly, only one member of the 32 DNA repair gene set reported by Polkinghorn et al. (2013) [
50],
HUS1, was detected in our 25 DNA repair gene set. The differences could be due to any number of biological variables associated with the cell lines or growth conditions in laboratories as our experiments were conducted using shorter R1881 treatment times.
It is possible that cell cycle changes pertaining to the observed androgen-induced G1 cell cycle arrest could contribute to the gene expression profiles detected in response to androgen - particularly because our analysis shows that all three cell lines displayed a strong androgen-regulated decrease in the E2F gene target hallmark (Fig.
3d). The E2F transcription factor family plays important roles as both activators and repressors of the cell cycle [
80]. Mechanistically, both transcription factors – AR and E2F1 – could contribute either cooperatively or independently to the changes in DNA damage response genes as has been shown previously for other androgen-regulated genes [
81].
Ultimately, our findings clearly support a role for androgen signaling in DNA repair gene expression in multiple cell lines, xenograft models, and in human tumor samples. As part of the analysis, we found that 60% of the DNA repair genes affected by androgen treatment in cell lines have AR binding sites based on AR ChIP-seq data from LNCaP and VCaP cells [
44]. The implication from our data is that androgen regulation of DNA repair gene expression could influence the response of prostate cancers to radiotherapy. But because there are both positive and negative effects on DNA repair gene expression, and there exists a complex interplay between DNA repair pathways, it is difficult to make simple predictions as to the biological outcome.
We also observed that DNA repair gene expression is changed during the transition from androgen-dependent to castrate-resistant cell growth. We examined publicly available microarray data from “HS-HR” xenograft pairs [
20] and found a total of 19 DNA repair genes were up- or down-regulated in at least four (of seven) models for CRPC. Using patient data, we found a set of 42 DNA repair genes were associated with metastasis. Some of these DNA repair genes have AR binding sites within 25 kb of the transcriptional start site or within the gene body itself, suggesting the regulation of a subset of these gene could be directly regulated through androgen signaling. The very small overlap between these sets of DNA repair genes might be explained by the fact each was generated from cells grown in vastly different milieu and selection pressures (cell culture, xenograft, human tumors). Genes enriched for specific pathways, namely cell growth and cell cycle, have been shown to be uniquely regulated in cell lines versus patient tumor samples [
82].
While the DNA repair machinery is widely appreciated for its role in correcting mutations generated during replication, and in response to various environmental insults, there is growing acceptance of its importance in transcription. The introduction of DNA breaks, for the purpose of resolving topological constraints [
59,
60,
63] and for expression of enhancer RNAs [
65], has been shown to be important for AR-dependent gene expression. Not surprisingly, androgen-dependent induction of these DNA breaks is accompanied by recruitment of DNA repair components, which occurs on a time scale of minutes. This mechanism might explain why certain DNA repair inhibitors reduce androgen-stimulated gene expression. Topoisomerase-mediated DNA breaks have been shown to be necessary and sufficient for transcription of genes that direct early events in neuronal differentiation [
83].
One of the DNA repair enzymes that was shown by the Rosenfeld group to be important for androgen-induced gene expression is MRE11, which is a component of the MRN complex [
65]. Mutation of
MRE11 was reported within the cohort of prostate cancer patients that showed a positive response to Olaparib [
6]. These findings led us to test whether the MRE11 inhibitor, mirin, can be used to inhibit AR-dependent transcription and prostate cancer growth. Mirin inhibition of growth in non-prostate cell types also required relatively high concentration of drug, with IC
50 values ranging from 12.5–100 μM [
67,
84‐
87]. This is comparable to the values we obtained in prostate cancer cells (Fig.
7e-f). We found that mirin inhibits transcription of multiple AR target genes in PC3-AR, LNCaP, and VCAP cells. The finding that androgen-induced expression of
FKBP5 was inhibited by mirin in all three cell lines argues that MRE11 complex function is critical for transcription in prostate cancer cells, which is consistent with the siRNA results published by another group [
65]. Androgen and mirin, used alone and in combination, did not have a noticeable effect on ATM levels, ATM activation, or H2AX phosphorylation. Thus, the mirin inhibition of AR-dependent transcription and cell growth is not accompanied by a global change in DNA damage signaling.