Identification of novel androgen-responsive genes by sequencing of LongSAGE libraries

Background The development and maintenance of the prostate is dependent on androgens and the androgen receptor. The androgen pathway continues to be important in prostate cancer. Here, we evaluated the transcriptome of prostate cancer cells in response to androgen using long serial analysis of gene expression (LongSAGE) libraries. Results There were 131 tags (87 genes) that displayed statistically significant (p ≤ 0.001) differences in expression in response to androgen. Many of the genes identified by LongSAGE (35/87) have not been previously reported to change expression in the direction or sense observed. In regulatory regions of the promoter and/or enhancer regions of some of these genes there are confirmed or potential androgen response elements (AREs). The expression trends of 24 novel genes were validated using quantitative real time-polymerase chain reaction (qRT-PCR). These genes were: ARL6IP5, BLVRB, C19orf48, C1orf122, C6orf66, CAMK2N1, CCNI, DERA, ERRFI1, GLUL, GOLPH3, HM13, HSP90B1, MANEA, NANS, NIPSNAP3A, SLC41A1, SOD1, SVIP, TAOK3, TCP1, TMEM66, USP33, and VTA1. The physiological relevance of these expression trends was evaluated in vivo using the LNCaP Hollow Fibre model. Novel androgen-responsive genes identified here participate in protein synthesis and trafficking, response to oxidative stress, transcription, proliferation, apoptosis, and differentiation. Conclusion These processes may represent the molecular mechanisms of androgen-dependency of the prostate. Genes that participate in these pathways may be targets for therapies or biomarkers of prostate cancer.


Background
Androgens mediate their effect through the androgen receptor (AR) and together they play integral roles in the development and maintenance of the prostate. In the absence of a functional androgen-axis during development, the prostate will fail to form [1]. The size of the pros-tate increases with the elevation of levels of androgens in males during puberty [2]. Androgens promote proliferation, differentiation, and survival of prostate cells [1]. Men that have used excess androgens in the form of anabolic steroids have a higher incidence of prostate cancer [3][4][5]. Association of prostate cancer with levels of androgens has also been reported in rodents [6,7]. Reduction of androgen in humans or dogs before puberty by castration is associated with decreased incidence of prostate cancer [8,9]. Castration of adult males causes apoptosis of prostatic epithelium, involution and reduction of the prostate [10][11][12]. Thus the prostate gland is an androgendependent organ where androgens are the predominant mitogenic stimulus [13]. The dependency of the prostate epithelium on androgens provides the underlying rationale for treating prostate cancer with chemical or surgical castration (androgen-deprivation) [14].
The AR is a ligand-activated transcription factor [15] that regulates transcription of genes that contain androgen response elements (AREs) in the upstream or downstream regulatory regions of the promoter and/or enhancer. Kallikrein 3 (KLK3) is an example of a gene that contains numerous functional AREs that the AR interacts with to increase transcription in response to androgens [16][17][18][19]. KLK3, also known as prostate-specific antigen (PSA), is the main tumor marker for prostate cancer and has been used clinically for 15 years [20]. Serum levels of PSA correlate with tumor volume [21]. However, as a screening and monitoring tool for prostate cancer, serum PSA levels are subject to false positives and false negatives [20].
Identification of the genes that change in expression in response to androgen in prostate cells is essential for the understanding of androgen-dependency of the normal prostate and the proliferation, survival, and hormonal progression of prostate cancer. There are several studies that have investigated genes that alter expression in response to a changing androgen-axis using SAGE [22][23][24]. Here we highlight several key differences in the current experimental design from previous studies: 1) a physiological concentration of metabolically stable androgen (R1881) was employed in vitro; 2) the transcriptome was catalogued using LongSAGE [25] opposed to (short)SAGE [26] because it generates lengthier tags allowing increased confidence in tag-to-gene mapping, and leaves fewer tags unmapped [25]; 3) the transcriptome of human prostate cancer cells was examined instead of murine cells [22]; 4) sequencing depth was increased by approximately 1.5-2 times more tags relative to other studies [23,24] to improve the potential for novel findings; 5) transcript expression was validated using an alternative assay as opposed to protein expression [24], and tens of novel genes were validated as opposed to only two [23]. Thus, we apply LongSAGE for the first time to create transcript libraries of prostate cancer cells maintained in the presence or absence of androgen. These libraries are publicly available at Gene Expression Omnibus. We describe 24 genes never before identified or validated to alter expression in response to androgen treatment. These genes were: ARL6IP5, BLVRB, C19orf48, C6orf66, CAMK2N1, CCNI, DERA, ERRFI1,  GLUL, GOLPH3, HM13, HSP90B1, MANEA, NANS,  NIPSNAP3A, SLC41A1, SOD1, SVIP, TAOK3, TCP1,  TMEM66, USP33, and VTA1. Statistically significant changes in expression of ARL6IP5, CAMK2N1, ERRFI1, HSP90B1, and TAOK3 in response to reduced levels of circulating androgens were measured using in vivo samples.

Summary of LongSAGE libraries
LongSAGE was employed to obtain quantitative gene expression profiles of human prostate cancer cells treated with or without synthetic androgen R1881. LNCaP human prostate cancer cells were chosen as the model cell line for evaluating androgen signaling because they respond to androgens, express a functional although mutated (T877A) AR, they can be grown in vitro as a monolayer or in vivo as a xenograft or in the Hollow Fiber model [27][28][29]. LNCaP cells have been used extensively in prostate cancer research. The time of 16 hours for treatment and concentration of R1881 (10 nM) were chosen based upon optimal induction of levels of KLK3 mRNA [30].
LongSAGE libraries were sequenced to a total of 121,760 (R1881) and 103,391 (vehicle) tags ( Table 1). The libraries were filtered on several levels to leave only useful tags for analysis. First, bad tags were removed if they contained at least one N-base call in the LongSAGE tag sequence. Notably, when bad tags were filtered the percentages of duplicate ditags in the R1881 and vehicle LongSAGE libraries were 6% and 5%, respectively. Early SAGE studies suggest duplicate ditags likely represent polymerase chain reaction (PCR) artifacts due to the low probability the same two tags will ligate together to form ditags [26]. However, with LongSAGE library sequencing and highly expressed transcripts, this probability becomes significant [31]. A recent study [32] suggests that discarding duplicate ditags in LongSAGE analysis may introduce a bias affecting the fold differences in tag expression between libraries for all tags observed at a frequency >(113-224)/100,000. Therefore, we opted to retain duplicate ditags. PHRED software was used to call bases for the sequencing of the LongSAGE tags [33,34]. PHRED has a small, but significant error rate in base-calls. To ascertain which tags potentially contained these erroneous basecalls, we calculated a tag sequence quality factor (QF) and probability [35]. The second line of filtering removed LongSAGE tags with probabilities less than 0.95 (QF < 95%). Linkers of known sequence were introduced into SAGE libraries as primers for amplifying ditags prior to concatenation [26]. These linker sequences were designed so they do not map to the human genome. At a low frequency, linkers ligate to themselves creating linkerderived tags (LDTs  t Tag count per 100,000 = (observed tag count/total tags in the library) × 100,000 u Transcript copies per cell w = (observed tag count/total tags in the library) × 500,000 v % Transcript abundance in cell w = (transcript copies per cell/500,000) × 100% w Calculation based on ~500,000 transcripts in a cell [36] α % of tags that map as transcription factors = (no. of genes with "transcription regulation acivity"/no. of genes with unambiguous sense mappings and GO terms) × 100% z Mapped unambigously sense toRefSeqand subjected to Gene Ontology (GO) analysis δ Tag types from each tag frequency class of R1881 and vehicle LongSAGE libraries were combined χ % of tags that map = (no. of genes with sense mappings/combined total tag types) × 1 00% β Mapped sense (incl. ambiguous) to RefSeq γ One tag was mapped sense using Ensembl gene ε % of tags significantly differentially expressed = (no. of significantly differently expressed tag types in class/combined total tag types in class) × 100% a Statistics according to the Audic and Claverie test statistic (p ≤ 0.001) ies per cell) and 5-9 times per 100,000 tags (25-45 transcript copies per cell) were the second and third most common groups of tag types, respectively. Generally, high frequency tags were less common. The majority of total tags in each LongSAGE library were derived from a few tag types detected between 10-99 times per 100,000 tags (50-495 transcript copies per cell).

Mapping distribution of LongSAGE tags
When mapped tags (v38 Ensembl) were clustered to amalgamate 1-off tags (see Methods, Gene Expression Analysis for a description) and tags that mapped ambiguously were removed, the tag types in the R1881 and vehicle Long-SAGE libraries represented 7,484 genes and 7,441 genes, respectively (Table 3). Tag types that mapped ambiguously constituted 13% (R1881 and vehicle), while 36% (R1881) and 35% (vehicle) of tag types did not map to the genome (Table 3). Due to the fact that these tags were clustered, the majority of the tags that did not map to the genome probably represent true unannotated transcripts rather than PCR/sequencing errors. Approximately 28% of tags in each LongSAGE library mapped to the opposite strand of known genes. These LongSAGE tags either represent transcription from previously undescribed coding regions or true antisense transcripts. Each LongSAGE library contained tags representing transcripts from 32% of the genes in the Ensembl gene database. This percentage is indicative of the depth of coverage of the transcriptome achieved with LongSAGE. Alternatively, this percentage indicates that one third of known Ensembl genes were expressed in LNCaP cells under these experimental conditions. This percentage is substantial when considering tag types from the Mouse Atlas Project (8.55 million total LongSAGE tags generated from 72 libraries of mouse development) mapped to 57% of the Ensembl transcript database [35]. Approximately 63% (R1881) and 61% (vehicle) of the genes that mapped to Ensembl's database were associated with more than one tag type to suggest that most gene expression was represented by transcript variants which is consistent with previous observations [35]. When the mapped LongSAGE tags (Reference Sequence, RefSeq; May 18, 2006) were clustered to amalgamate 1-off tags and tags that mapped ambiguously were removed, 53% of tags mapped solely to known exons, 9% solely to known introns (novel transcript vari-ants), and 38% to intergenic regions (novel genes or transcript variants).
The two most abundant tag types in the LongSAGE libraries were shared by both libraries. The first highly abundant LongSAGE tag mapped to human mitochondrial NADH ubiquinone oxidoreductase chain 4. This gene is also highly expressed in other human tissues (i.e., cardiac tissue; SAGE Genie, http://cgap.nci.nih.gov/SAGE). The protein product of this gene transfers electrons from NADH to ubiquinone to generate adenosine triphosphate as metabolic energy. Using the Ensembl database, the second most abundant LongSAGE tag mapped to a non-coding gene of human mitochondria. In contrast to the higher abundance classes, the lower abundance classes were enriched for LongSAGE tags that mapped to genes with functions in regulating transcription ( Table 2). This is particularly significant because the percentages of Long-SAGE tags that mapped to the genome in the lower abundance class were reduced compared to the higher abundance classes ( Table 2). Together this implies that the number of tags that map to genes with a function in transcription may be underestimated, as low abundance tags may be underrepresented.

Differential gene expression
Venn analysis identified that 36% and 38% of tag types were exclusive to the R1881 or vehicle LongSAGE libraries, respectively ( Figure 1). The unique expression of tag types indicates differential expression depending upon androgen stimulation. The biological relevance of this differential expression is complicated by the fact that 85% (R1881) and 88% (vehicle) of these exclusive LongSAGE tags were singletons. Consistent with our observation that low abundance tags did not map as readily to the genome, the mutually exclusive tags also did not map as readily as tags shared between both libraries. Only 17% and 15% of tags exclusive to R1881 and vehicle LongSAGE libraries, respectively, mapped unambiguously sense to RefSeq, in contrast to 39% of shared tags. We therefore, concentrated on genes for which the tag abundance allowed for the determination of statistically significant changes in transcript abundance.  Figure 2). 891 tags were differentially expressed (p ≤ 0.05) between the two LongSAGE libraries ( Figure 2 and Table 4). Long-SAGE tags statistically (p ≤ 0.001) differentially represented between the libraries were enriched in the higher abundance classes compared to the lower abundance classes (Table 2). Additionally, 90% of the LongSAGE tags were statistically (p ≤ 0.001) differentially represented between the libraries with ≥ 2-fold differences, compared to only 17% of tags with p-values greater than 0.001 (p > 0.001).
A stringent p-value cutoff (p ≤ 0.001), not corrected for multiple tests, was employed prior to validation of changes in expression of a gene in response to androgen. LongSAGE tags that were differentially expressed, but mapped ambiguously to more than one gene, and/or dif-fered by less than 2-fold between the treatment groups, were excluded from analysis. Application of these criteria reduced the LongSAGE tags from 131 to 93. These 93 tags represented 87 genes. Analysis of differentially expressed LongSAGE tags revealed that 54 LongSAGE tags that mapped to 52 genes were previously known to change in expression in the direction observed in response to androgen in prostate cancer cells. Of these, the expression of 41 genes increased as expected, including the well-known androgen-regulated gene, KLK3 ( Table 5). The expression of 11 genes decreased in response to androgen and were consistent with previous reports (Table 6). Genes previously not reported to alter expression in response to androgen in prostate cancer cells were represented by 39 LongSAGE tags. These tags represented the expression of 20 genes that were increased, excluding mappings to noncoding and intergenic regions, (Table 7), and expression of 15 genes that was decreased (Table 8) in response to androgen. The 93 tags were represented by 87 genes because one tag did not map to the human genome (Table  7) and two tags mapped to intergenic regions of the human mitochondrial genome (Tables 7 and 8). Three genes were represented twice in the tables (CAMK2N1, PPAP2A, and SORD). One gene, KRT8, was categorized in both the known and not previously known categories due to the sense of the mapping (Tables 5 and 8).
Interestingly some antisense tags were identified as differentially expressed in response to androgen. Antisense to NKX3-1 is of particular note. Transcription of this gene is Figure 1 Relationship between LongSAGE library compositions. The Venn Diagram shows the tag types and genes exclusive to, and shared by each LongSAGE library, R1881 and vehicle. Tags were mapped unambiguously sense to Ref-Seq transcripts and redundant mappings were removed. Singletons are tags counted only once in each library, but may be common to both libraries.    however, this remains to be determined.

Validation of changes in gene expression in response to androgen
Quantitative real time-polymerase chain reaction (qRT-PCR) was used to validate changes in gene expression in response to androgen of 39 (13 known; 26 novel) of the 87 total genes identified by LongSAGE. Of the 35 genes previously not reported to change expression in response to androgens in prostate cancer cells, only 26 were quantified by qRT-PCR, because technical limitations and gaps in the transcriptome databases prevented the analysis of 9 genes. That is, specific qRT-PCR primers could not be designed due to repetition in the genome, or because the tag mapped to an unannotated transcript variant. There were 24 of the 26 (92%) novel genes that displayed statistically significant differential expression in response to androgen as measured by qRT-PCR ( Figure 3A). BLVRB, C19orf48, C1orf122, ERRFI1, GLUL, GOLPH3, HM13, HSP90B1, NANS, SLC41A1, TAOK3, TCP1, TMEM66, and USP33 all increased levels of expression in response to androgen, while ARL6IP5, C6orf66, CAMK2N1, CCNI,   Figure 3B).

Known or potential AREs in the regulatory regions of androgen-regulated genes
AR directly regulates transcription in response to androgen by binding to AREs in the promoter and/or enhancer regions of target genes. ChIP-chip database mining for suggested AREs combined with a literature search for known AREs revealed some of the genes that alter expression in response to androgen do contain AREs ( found eight genes from our gene list to contain potential AREs ( Table 9). Identification of potential AREs in the regulatory regions of the newly identified genes that alter expression in response to androgen (BLVRB, C19orf48, HM13, SOD1) may be directly regulated by AR.

Cell-type specificity of gene expression
To determine if expression of candidate genes was unique to LNCaP cells, we assayed for constitutive levels of expression of 18 known and novel candidate genes in prostate cancer cell lines DU145[53] and PC-3[54] using qRT-PCR ( Figure 4). Genes chosen included those that both increased (ADAMTS1, CAPNS1, CENPN, CREB3L4, ERRFI1, FKBP5, HSP90B1, KLK3, LRIG1, NCAPD3, PAK1IP1, and TAOK3) and decreased expression in response to androgen (ARL6IP5, CAMK2N1, CCNI, CXCR7, PRKACB and ST7). No obvious trends were observed depending on whether expression of the genes increased, or decreased, in response to androgen. All genes tested, except ERRFI1, were expressed at a lower level in PC-3 and DU145 cells relative to LNCaP cells. This suggests that the majority of genes that alter levels of expression in response to androgen were enriched in LNCaP cells relative to PC-3 and DU145 cells. These data are consistent with both DU145 and PC3 cells being androgen-insensitive and lacking a functional AR[53,54].

In vivo changes in gene expression in response to androgen-deprivation
The LNCaP Hollow Fibre model combined with qRT-PCR was employed to capture in vivo gene expression representative of physiological levels and castrated levels of androgen ( Figure 5). We expected that the genes that had increased levels of expression in vitro in response to androgens, would decrease expression in vivo in response to castration (androgen-deprivation). Conversely, we expected that the genes that had decreased levels of expression in vitro in response to androgens, would increase expression in vivo in response to castration. These in vivo results would be consistent with androgen-responsiveness of the candidate genes. Of the candidate genes examined, 13 of 16 genes showed significant changes in gene expression in response to androgen-deprivation ( Figure 5). As anticipated, expression of ARL6IP5, CAMK2N1, CXCR7, and ST7 increased, while CENPN, CREB3L4, ERRFI1, FKBP5, KLK3, LRIG1, NCAPD3, PAK1IP1, and TAOK3 decreased a Statistics according to the Audic and Claverie test statistic (p ≤ 0.001) b ND, not detected c ND tags were assigned a value of 1 when calculating fold change d Appropriate significant figures are displayed e Gene family, but not this family member, previously described to change expression in response to androgens g Protein known to change expression in reponse to androgens h Gene known to change expression in response to androgens, but in the opposite direction as reported here i Gene known to change expression in response to androgens in cells other than prostate k Tag has a single base pair mutation, insertion, or deletion with respect to gene map mTag maps to the strand opposite of the gene n Ambiguously mapped tags and tags with a fold change less than 2-fold have been excluded from the table p NC_001807, refers to the complete genome of mitochondria in humans All mitochondrial genes in the RefSeq database are assigned the same accession number by NCBI q N/A, there is no HGNC approved gene symbol or description for this tag t Tag count per 100,000 = (observed tag count/total tags in the library) × 100,000 φ In cases where a tag mapped to >1 transcript variant of the same gene the RefSeq accession number for transcript variant 1 was displayed * Gene further characterized in this paper

Conclusion
Androgens are essential for the growth, development and maintenance of the prostate. Here, we created LongSAGE libraries to obtain quantitative gene expression profiles of LNCaP human prostate cancer cells treated with, or with-Androgen regulation of genes as measured by qRT-PCR Figure 3 Androgen regulation of genes as measured by qRT-PCR. A Candidate genes not previously implicated to change expression in response to androgens in prostate cancer cells, and B Genes known to change levels of expression in response to androgens. LNCaP cells were treated for 16 hours prior to harvesting RNA, and analysing mRNA levels by qRT-PCR. Foldchange was calculated by normalizing the mean normalized expression (MNE) of transcripts in R1881-treated cells to the mock vehicle-treated cells. In doing this, the vehicle treatment fold-change became one and standard deviation (SD) zero. Error bars represent ± SD for biological sextuplets.
[*] Asterisk indicates significant differential gene expression according to the Two-Sample Student's T-test (p ≤ 0.05) for unequal variance.   out, androgen and revealed the following: 1) 33,385 tag types in the R1881 LongSAGE library and 31,764 tag types in the vehicle LongSAGE library; 2) the majority (64% to 67%) of tag types in each LongSAGE library were singletons which may represent very low abundance transcripts (≤ 5 transcript copies per cell); 3); when mapped tags were clustered and ambiguous mappings were removed, the tag types in the R1881 and vehicle LongSAGE libraries represented 7,484 genes and 7,441 genes, respectively; 4) 53% of tags mapped solely to known exons, 9% solely to known introns (novel transcript variants), and 38% to intergenic regions (novel genes or transcript variants); 5) the most highly abundant LongSAGE tag mapped to human mitochondrial NADH ubiquinone oxidoreduct-ase chain 4 involved in metabolic energy; 6) the lower abundance classes were enriched for genes with functions in regulating transcription; 7) 87 genes were differentially expressed by two-fold (p ≤ 0.001) in response to androgen representing 0.34% of the total tag types (131 differentially expressed tag types/38,574 total tag types); 8) some of these genes have confirmed or potential AREs; 9) novel androgen regulated genes (direct or indirect) identified and validated were ARL6IP5, BLVRB, C19orf48, C1orf122 ,  C6orf66, CAMK2N1, CCNI, DERA, ERRFI1, GLUL,  GOLPH3,  HM13,  HSP90B1,  MANEA,  NANS,  NIPSNAP3A, SLC41A1, SOD1, SVIP, TAOK3, TCP1,  TMEM66, USP33, and VTA1; 9) expression of ADAMTS1, ARL6IP5, CAMK2N1, CAPNS1, CENPN, CREB3L4, CCNI,  Asterisks indicate the significant differential gene expression in each cell line compared to LNCaP cells according to the Two-Sample Student's T-test (p ≤ 0.05) for equal (unpaired) or unequal variance as determined appropriate with the F-test. , and TAOK3 were measured in vivo in response to androgen-deprivation. The products of these genes are involved in amino acid and protein synthesis, cofactor transport, protein trafficking, response to oxidative stress, as well as signaling pathways that regulate gene expression, proliferation, apoptosis, and differentiation.
These genes are potentially critical for the function and maintenance of the prostate and represent targets for clinical intervention.   were retained for analysis. Sequence data were filtered for bad tags (tags with one N-base call) and linkerderived tags (artifact tags). Only LongSAGE tags with a sequence quality factor (QF) greater than 95% were included in analysis [35]. Where indicated, a clustering algorithm was used to amalgamate 1-off tags (tags one bp incorrect from a complete map to a transcript) with likely 'parent' tags to improve the mapping capability of Long-SAGE tags by apparently reducing PCR/sequencing errors [35]. This clustering algorithm altered the number of tag types (i.e., species) without changing the total number of tags. In instances where clustering was used, the 95% QF cutoff was not. To filter data for candidate transcript validation, a p-value cutoff (p ≤ 0.001) was employed according to the Audic and Claverie test statistic [40]. The Audic and Claverie statistical method was used to identify differentially expressed tags between LongSAGE libraries because the method takes into account the sizes of the libraries and tag counts. Long-SAGE tags that mapped ambiguously to more than one gene, and tags that differed by less than 2-fold were excluded from the candidate list. LongSAGE tags were mapped to reference sequence (RefSeq; May 30 th , 2005) and Ensembl Gene (v31.35d), unless otherwise stated.

Quantitative real-time polymerase chain reaction
qRT-PCR was performed on TRIZOL-extracted RNA from LNCaP (serum-starved ± R1881 or the exception in Figure  4 in 10% serum), DU145 (10% serum) and PC-3 (5% serum) cells maintained in vitro, and LNCaP cells maintained in the in vivo Hollow Fibre model [29] (see below). Contaminating genomic DNA was removed from in vitro RNA samples using DNA-free or TURBO DNA-free (Ambion, Austin, TX, USA). Input RNA (1 μg) was reverse transcribed with SuperScript III First Strand Synthesis kit (Invitrogen). A 10 μL qRT-PCR reaction included 1 μl of template cDNA (0.1 μL for limited LNCaP Hollow Fibre samples), 1× Platinum SYBR Green qPCR SuperMix-UDG with ROX (Invitrogen) and 0.3 μM each of forward and reverse intron-spanning primers that produce products between 85-115 bp in size (see Additional file 1 for primer sequences). qRT-PCR reactions were cycled as follows in a 7900 HT Sequence Detection System (Applied Biosystems): 50°C for 2 min, 95°C for 2 min, (95°C for 0.5 min, 55-56°C for 0.3-0.5 min, and 72°C for 0.5 min) for 40-45 cycles, 95°C for 0.25 min, 60°C for 0.25 min, and 95°C for 0.25 min. All qRT-PCR reactions were performed in technical triplicates for each of at least three biological replicates. cDNAs (from different conditions) and genes [target and reference (glyceraldehyde-3-phosphate, GAPDH)] to be directly compared were assayed in the same instrument run. No-template reactions (negative controls) were run for each gene to ensure that DNA had not contaminated the qRT-PCR reactions. Only qRT-PCR data with single-peak dissociation curves were included in this analysis. Efficiency checks were performed for each primer pair in each cell line. PCR products were sequenced to verify the identity of quantified transcripts. The two-tailed, two-sample Student's T-tests were performed to identify significant differences in transcript expression. The F-test was used to identify unequal variance among samples to be compared.

LNCaP Hollow Fibre model Animals
Five-week-old male athymic BALB/c Nude mice were obtained from Taconic Farms (Hudson, NY, United States of America) and kept in the British Columbia Cancer Research Centre (Vancouver, BC, Canada). Mice were maintained on a Harlan/Teklad irradiated diet with a constant supply of autoclaved water and housed in cages (three animals/cage) at 21°C ± 3°C with light/dark cycling (light between 6 AM and 6 PM). All animal experiments were performed according to a protocol approved by the Committee on Animal Care of the University of British Columbia.

Hollow fibre model
Polyvinylidene difluoride hollow fibres (M r 500,000 molecular weight cutoff; 1-mm internal diameter; Spectrum Laboratories, Rancho Dominguez, CA, USA) were prepared and implanted as previously described [29]. Briefly, LNCaP human prostate cancer cells (3 × 10 7 cells) at passage 47 (provided by Dr. L.W.K. Chung at the Emory University School of Medicine, Atlanta, GA, USA) were injected into hollow fibres. The fibres were sealed and subcutaneously (s.c.) implanted into mice. Seven days post fibre implantation (day zero), mice were either castrated or left intact as controls. Blood was drawn via the tail vein each week to measure serum KLK3 levels to monitor the response to castration. Serum KLK3 levels were determined by enzymatic immunoassay kit (Abbott Laboratories, Abbott Park, IL, USA). Bundles of fibres were removed at day zero (Pre-Cx; four fibres) and day 10 (Cx; four fibres). Total RNA was isolated immediately from cells harvested from the fibres. Compromised fibres that were contaminated with mouse cells, as indicated by an infiltration of red blood cells that was determined by visual inspection, were not used in this study.