Genome-wide CRISPR screen identifies HNRNPL as a prostate cancer dependency regulating RNA splicing

Significance

Alternative RNA splicing and the spliceosome machinery have been implicated in cancer progression. A genome-wide CRISPR screen identified the RNA processing factor heterogeneous nuclear ribonucleoprotein L (HNRNPL) as required for prostate cancer growth by regulating alternative RNA splicing and circular RNA formation. HNRNPL and its RNA clients are overexpressed during prostate cancer progression, supporting their potential role as therapeutic targets.

Abstract

Alternative RNA splicing plays an important role in cancer. To determine which factors involved in RNA processing are essential in prostate cancer, we performed a genome-wide CRISPR/Cas9 knockout screen to identify the genes that are required for prostate cancer growth. Functional annotation defined a set of essential spliceosome and RNA binding protein (RBP) genes, including most notably heterogeneous nuclear ribonucleoprotein L (HNRNPL). We defined the HNRNPL-bound RNA landscape by RNA immunoprecipitation coupled with next-generation sequencing and linked these RBP–RNA interactions to changes in RNA processing. HNRNPL directly regulates the alternative splicing of a set of RNAs, including those encoding the androgen receptor, the key lineage-specific prostate cancer oncogene. HNRNPL also regulates circular RNA formation via back splicing. Importantly, both HNRNPL and its RNA targets are aberrantly expressed in human prostate tumors, supporting their clinical relevance. Collectively, our data reveal HNRNPL and its RNA clients as players in prostate cancer growth and potential therapeutic targets.

Prostate cancer is among the most prevalent adult malignancies in developed countries. The principal treatment for prostate cancer once it is no longer amenable to surgery or radiation treatment is androgen deprivation therapy, which targets androgen or androgen receptor (AR) signaling. However, resistance to androgen deprivation therapy often develops and leads to a state termed “castration-resistant prostate cancer,” which still lacks an effective cure (1–3). Therefore, significant efforts have been devoted to better understand the mechanism of oncogenesis and to develop additional effective therapeutics targeting pivotal oncogenes, cancer-related signal transduction pathways, and epigenetic regulators (4, 5).

Alternative RNA splicing is a fundamental cellular process by which a single gene can give rise to multiple different transcripts and proteins. This process is tightly regulated by core spliceosomes and other splicing factors, such as the serine/arginine-rich family of proteins and heterogeneous nuclear ribonucleoproteins (hnRNPs) (6, 7). Multiple studies indicate that deregulation of alternative splicing is implicated in cancer progression and that the splicing machinery may be targeted therapeutically (8–10). In addition to RNA splicing, the physical interactions between RNAs and RNA binding proteins (RBPs) underlie multiple RNA processing steps, such as capping, polyadenylation, transport, localization, modification, and translation, thereby regulating many aspects of RNA fate (11). Which RBPs and their related RNA processing steps are functionally important, especially in prostate cancer, remains elusive.

The recent implementation of the clustered regularly interspaced short palindromic repeats (CRISPR)-associated nuclease Cas9 genome editing system has proved effective in high-throughput loss-of-function screens (12–14). Compared with RNA interfering (RNAi)-based gene knockdown, CRISPR/Cas9-based gene knockout confers a more thorough deletion of target gene expression and has less off-target effects when the guide RNA is appropriately designed (15, 16). The power of the pooled CRISPR screen has been demonstrated by several studies investigating the genes involved in drug and toxin resistance (13, 17), cancer metastasis (18), and immune response (19).

Here, by a genome-wide CRISPR screen and in depth mechanistic studies, we sought to systematically identify functional RBPs or RNA processing factors that are essential for prostate cancer growth and underlie pivotal cancer-related RNA processing steps, especially RNA splicing, thus potentially enabling the development of novel cancer therapeutics targeting RBPs or RBP–RNA interactions.

Results

Genome-Wide CRISPR/Cas9 Knockout Screen in Prostate Cancer Cells.

To systematically identify the functional RBPs or RNA processing factors in prostate cancer, we performed a pooled genome-wide CRISPR/Cas9 knockout screen in prostate cancer LNCaP cells using the GeCKO v2 lentiviral single guide RNA (sgRNA) library (20) (Fig. 1A). This library contains 123,411 unique sgRNAs targeting 19,050 genes, with 6 sgRNAs per gene. After selection with puromycin, half of the infected cells were harvested as a day_0 control sample and the rest of the cells were continuously cultured for an additional two weeks as the day_14 sample. The sgRNAs incorporated in the cells were then amplified from genomic DNA and subjected to massively parallel sequencing for quantification. We defined the negatively selected essential genes as those with significantly depleted sgRNAs at day 14 and positively selected genes as those with significantly enriched sgRNAs at day 14 (Dataset S1). MAGeCK and MAGeCK-VISPR are statistical algorithms we previously developed for CRISPR screen analyses (21, 22). Whereas MAGeCK was developed to call genes from a single CRISPR screen experiment, MAGeCK-VISPR was developed to compare the gRNA abundance across many conditions. The magnitude of negative selection for the top 1,000 essential genes in LNCaP cells is comparable to that of known essential ribosomal protein genes (Fig. S1A). We thus chose these top 1,000 genes as the LNCaP essential gene set for the downstream analysis. By comparing this list of essential genes to publically available data for 10 other cell lines, we found that ∼one-third of the LNCaP essential genes are broadly essential across multiple cell types (as pan-essential genes), whereas the remaining genes were either LNCaP-specific or essential only in a subset of cell types (Fig. 1B). The most enriched functional categories for the pan-essential genes fall into protein biosysthesis, ribonucleoprotein, ribosome, proteasome, and cell cycle that are all core functional components or processes in the cell (Fig. S1B). In terms of enriched pathways, LNCaP essential genes are enriched for AR and PI3K pathways that are known to play pivotal signaling roles in LNCaP cells, supporting the relevance of our screen (Fig. S1C). In addition, many known functional genes in prostate cancer were under significant negative selection, including STAT3, PIAS1, PIK3C2A, AR, and MYC (Fig. 1C) (23–25). Overall, these top 1,000 essential genes established an interconnected network highlighting multiple key signaling molecules, complexes, or processes, including spliceosome and ribonucleoproteins (Fig. 1D). Furthermore, we investigated the gene copy number effects on CRISPR screen performance and confirmed that genes in genomic amplified regions tend to exhibit more negative selection as revealed by recent studies (16, 26) (Fig. S1D and Dataset S1). These potential false-positive hits from copy number-increased regions were removed from the downstream functional analysis.

Fig. 1.

Genome-wide CRISPR/Cas9 screen in prostate cancer. (A) Workflow of the CRISPR/Cas9 screen in LNCaP cells cultured in the presence of 10 nM dihydrotestosterone (DHT) with charcoal/dextran-treated FBS-containing medium. (B) Multiple CRISPR screening datasets (12, 15, 49) reveal the functions of LNCaP essential genes across different cell types. The β-essentiality scores of top 1,000 LNCaP essential genes, calculated by MAGeCK-VISPR, are shown. Negative (or positive) β-score indicates the corresponding gene is undergone negative (or positive) selection. The less the β-score, the more essential the corresponding gene. These genes consist of two categories: pan-essential genes (genes that are also essential in more than half of the other cell types) and LNCaP-specific essential genes. (C) CRISPR screen of essential genes in LNCaP cells. Genes were rank-ordered by robust rank aggregation (RRA) scores calculated by MAGeCK (21), where a smaller RRA score indicates more essentiality. HNRNPL and known LNCaP functional genes are highlighted. (D) A network view of the top 1,000 LNCaP essential genes. Nodes represent genes, and an edge connecting two genes if both are in the same pathway. AR and some known LNCaP oncogenes are highlighted, and some major gene clusters are also marked using different colors. The pathway information is extracted from GeneMANIA database (50).

Fig. S1.

Genome-wide CRISPR screen in prostate cancer cells. (A) Fold-change of sgRNA abundance between day 14 and day 0 samples for the indicated groups of genes: ribosomal essential genes (Ribo_essential); top 1,000 essential genes (1k_essential); top 1,000 positively selected genes (1k_positive) and nonselective genes. *P < 6.75e-11. (B) The most significantly enriched functional categories of pan-essential genes calculated by DAVID functional annotation tools. (C) The most significantly enriched pathways for LNCaP essential genes. The pathway gene sets were extracted from MSigDB (software.broadinstitute.org/gsea/msigdb). (D) The distributions of β-essentiality scores of all genes, grouped by the copy number status (measured in log₂ ratio) of the gene in LNCaP cells. Negative (or positive) β-score indicates the corresponding gene is undergone negative (or positive) selection. The lower the β-score, the more essential of the corresponding gene. *P = 0.035 by Wilcox rank sum test. Copy number variation (CNV) data were downloaded from The Cancer Cell Line Encyclopedia project (https://portals.broadinstitute.org/ccle/home).

HNRNPL Is Required for Prostate Cancer Cell Growth.

To explore which factors might play a role in RNA processing, we examined the RBPs and RNA splicing factors that were essential in the screen. Interestingly, several HNRNP family members, such as HNRNPL, HNRNPC, and FUS, are among the top essential RBPs (Dataset S1). Notably, HNRNPL was the top HNRNP essential gene in LNCaP cells and comparable to the known LNCaP essential genes identified by CRISPR screen, such as AR (Fig. 1C and Dataset S1). To cross-validate and more carefully examine the HNRNP family genes, we additionally used a targeted short interfering RNA (siRNA) approach to knock down each HNRNP gene with two independent and effective siRNAs, and examined the effects on cell growth after 6 d of siRNA exposure. Among all of the HNRNP genes tested, HNRNPK and HNRNPL showed the most dramatic and consistent cell growth reduction upon knockdown, followed by HNRNPC (Fig. 2A and Fig. S2A). HNRNPK was not identified as essential in the CRISPR screen, which might be because of differences between the CRISPR and RNAi technologies, such as inefficient sgRNAs or siRNA off targets.

Fig. 2.

HNRNPL is a functional RBP in prostate cancer. (A) Validation of essentiality of HNRNPs using siRNA oligos in LNCaP cells. LNCaP cells cultured in full medium were transfected with corresponding siRNAs and cell growth was determined by cell counting after 6 d of transfection. Each HNRNP was targeted by two independent and effective siRNA oligos (siRNA-a and siRNA-b). Data are shown as relative cell growth by comparison with control knockdown cells, mean ± SD, n = 3. Functional HNRNPs, denoted with an asterisk, are defined as exhibiting less than 0.5 of relative cell growth by two independent knockdowns. (B) Cell growth validation upon HNRNPL and HNRNPK knockdown using an additional two independent siRNA oligos (siRNA-1 and siRNA-2) in three cell lines: normal-like RWPE-1 prostate cells, LNCaP androgen-dependent prostate cancer cells, and CWR22Rv1 castration-resistant prostate cancer cells. Cells were counted after 6 d of siRNA transfection and data are shown as mean ± SD, n = 3. Knockdown efficiency was determined by Western blot analysis.

Fig. S2.

HNRNPL is an essential RBP in prostate cancer. (A) RT-qPCR analysis of corresponding HNRNP gene expression upon two independent HNRNP knockdowns (siRNA-a and siRNA-b). Data are shown as mean ± SEM, n = 3. (B) Soft agar assay of LNCaP cells upon HNRNPL or HNRNPK knockdown. (C) Cell-growth effect of HNRNPL or HNRNPK knockdown in two AR-null prostate cancer cell lines DU145 and PC3. (D) Superenhancer-associated HNRNP genes across 21 cell lines. The red box indicates the gene associated with a superenhancer in the respective cell line. The catalog of the superenhancer in different human cell types was obtained from ref. 27. A gene is considered to be associated with a superenhancer if its transcription start site is within 10-kb from the superenhancer.

To further minimize the potential siRNAs off-target effects for HNRNPK and HNRNPL, we used two additional independent siRNAs targeting either HNRNPK or HNRNPL, and examined the cell-growth effect in several prostate cancer cell line models. In both the androgen-dependent LNCaP cells and the castration-resistant CWR22RV1 cells, as well as the AR-negative prostate cancer DU145 and PC3 cell lines, knockdown of either HNRNPL or HNRNPK could diminish cell growth (Fig. 2B and Fig. S2 B and C). Interestingly, in a normal-like prostate epithelial cell line, RWPE-1, HNRNPL is not required for cell growth, whereas HNRNPK remains essential (Fig. 2B), indicating that HNRNPL, but not HNRNPK, may have cancer-specific functions. Moreover, recent studies suggested that superenhancer-associated genes tend to be functionally important in the control of cell identity and disease (27). HNRNPL and PTBP1 were the only members of the HNRNP family categorized as superenhancer-associated genes in LNCaP cells (Fig. S2D). Taken together, our CRISPR screen, siRNA validation, and superenhancer analysis identified HNRNPL, among the other RBPs, as worthy of further study.

RNA Binding Landscape of HNRNPL by RNA Immunoprecipitation Coupled with Next-Generation Sequencing.

As a classic RBP, HNRNPL functions through protein–RNA interaction. To characterize the genome-wide landscape of HNRNPL-associated RNAs, we used RNA immunoprecipitation coupled with next-generation sequencing (RIP-seq) to map the HNRNPL–RNA interactome in LNCaP cells. Similar to ChIP for analyzing protein–DNA interactions, the RIP approach uses a specific antibody against the protein of interest to pull down formaldehyde cross-linked RNAs. Biological replicates of HNRNPL RIP-seq identified 6,892 consensus binding sites across 1,549 genes (Fig. 3A and Dataset S2). The reproducibility of our RIP-seq data was evidenced by the correlation of read count across 10-kb windows between replicates (r = 0.98) (Fig. 3B). We further categorized the distribution of the binding sites across different genomic elements and found that HNRNPL-associated RNA peaks are mostly enriched in introns and 3′UTRs (Fig. 3C and Fig. S3A). De novo motif analysis showed that the most enriched binding motif in LNCaP cells is a CA-repeat or CA-enriched pattern (Fig. 3D), consistent with previous studies of HNRNPL–RNA interactions in other cell types (28, 29).

Fig. 3.

The genome-wide landscape of HNRNPL binding sites on RNA. (A) Venn diagram of HNRNPL RIP-seq peaks from two biological replicates. LNCaP cells cultured in charcoal/dextran-treated FBS were treated with 10 nM androgen DHT for 4 h and harvested for RIP-seq analysis. The peaks from two biological replicates were compared. (B) Scatter plot and correlation coefficience of the RIP-seq read count within a 10-kb window between two replicates. (C) HNRNPL RIP-seq peaks were enriched in introns and 3′UTR region. All RIP-seq peaks were categorized according to the distribution on different genomic elements and compared with the genomic background. (D) De novo motif analysis identifying CA-repeat motif as the only enriched motif within the top HNRNPL RIP-seq peaks. (E) Representative HNRNPL RIP-seq peaks were shown as track signals in an integrative genomic viewer. Peak area is highlighted in light green and peak region is denoted as red rectangle. RPS28 served as a negative control for HNRNPL enrichment. (F) RIP-qPCR analysis of HNRNPL binding sites as shown in E. The left side peak of ROR2 in E was chosen for RIP-qPCR validation. IgG RIP served as negative RIP control. RPS28 served as negative site for RIP enrichment. Data were shown as mean ± SEM, n = 3.

Fig. S3.

RIP-seq analysis of HNRNPL and AR is a downstream target of HNRNPL. (A) Genome browser representation of previously reported HNRNPL binding (NOS3 and TCF3), 3′UTR binding (PORCN), and exon binding (TRIOBP) events. (B) RNA pulldown validation of several indicated HNRNPL-bound regions. The HNRNPL binding peak regions from indicated genes were PCR-amplified and in vitro transcribed to produce either the sense or antisense strand of corresponding RNAs with biotin labeling. The biotin-labeled RNA fragments were then incubated with LNCaP cell lysate and pulled down by streptavidin beads. RNA-associated proteins were resolved by Western blot analysis. EGFP and RPS28 served as negative control for HNRNPL binding. (C) Genome browser representation of HNRNPL binding and the differentially spliced cryptic exon (exon 2b) over AR loci from RIP-seq and RNAseq data. (D) Overexpression of the full-length AR only partially rescues the cell growth effect upon HNRNPL knockdown. (Left) LNCaP cells with stably transfected full-length AR and parental cells were used to determine the cell-growth effect upon HNRNPL knockdown. Cells were counted after 6 d of siRNA transfection. Western blot was performed 3 d after siRNA transfection. (Right) LNCaP cells were infected with lentivirus expressing doxycycline-inducible full length AR to determine the cell-growth effect upon HNRNPL knockdown; 500 ng/mL doxycycline (Dox) was used and cell counting was performed after 6 d of induction. Western blot was done 3 d postinduction.

To further validate the RIP-seq peaks, we used two independent monoclonal HNRNPL antibodies (4D11 and D5) to perform RIP-qPCR analysis on several representative RIP-seq peaks with differential binding strengths. Although HNRNPL does not bind to the negative control RPS28 transcript, it has significant but differential binding for CTBP1, ROR2, and STX3 transcripts (Fig. 3 E and F). In addition, we performed RNA pulldown assays to validate these peak regions. Biotin-labeled peak-region RNA was in vitro transcribed and incubated with cell lysate. Streptavidin beads then captured the RNA–protein complex and interacting proteins were resolved by Western blot analysis. Only the sense strand of peak RNAs, but not the antisense strand, could associate with HNRNPL, further proving the sequence specificity of HNRNPL–RNA association (Fig. S3B). Taking these data together, using RIP-seq we identified a specific set of HNRNPL-associated RNAs in prostate cancer with high confidence and precision, suggesting that these may underlie the mechanism of HNRNPL function.

HNRNPL Regulates Alternative Splicing in Prostate Cancer Cells.

The significant intron-binding pattern implied that HNRNPL is primarily bound to pre-mRNA and might directly regulate alternative splicing, consistent with a previously reported function of HNRNPL (29, 30). To gain a comprehensive view of HNRNPL-dependent alternative splicing events in prostate cancer, we transfected either control or HNRNPL-targeting siRNAs in LNCaP cells and applied paired-end sequencing on the polyadenylated transcriptome on three biological replicates. Alternative splicing events were analyzed by the rMATS algorithm (31). In total, 206 alternative splicing events were identified as HNRNPL-dependent, with the majority falling in the skipped exon category (Fig. 4A and Dataset S3). HNRNPL binding is significantly associated with these splicing changes (Fig. 4B), supporting a direct effect of HNRNPL–RNA interaction on alternative splicing.

Fig. 4.

HNRNPL directly regulates alternative splicing in prostate cancer. (A) Pie chart showing different classes of HNRNPL-dependent alternative splicing events. (B) HNRNPL binding is significantly associated with HNRNPL-dependent alternative splicing. Fraction of genes with HNRNPL binding from differential splicing genes was compared with that from all genes. (C) Genome browser representation of HNRNPL binding over AR loci from RIP-seq data. (D) RT-qPCR confirmation of HNRNPL-dependent alternative splicing of AR. Both two alternatively spliced isoforms and pre-mRNA of AR were examined using corresponding PCR primers by RT-qPCR analysis upon HNRNPL knockdown. Data were shown as mean ± SEM, n = 3, *P < 0.05. (E) RIP-qPCR confirmation of HNRNPL binding on AR RNA. Data were shown as mean ± SEM, n = 3. (F) Western blot analysis of full-length AR protein level upon HNRNPL knockdown. GAPDH served as loading control.

Interestingly, the AR pre-mRNA is significantly bound by HNRNPL and an HNRNPL-dependent alternative splicing event was identified in the AR transcript (Fig. 4C, Fig. S3C, and Dataset S3). A cryptic exon 2b was identified as alternatively spliced between constitutive exon 2 and exon 3 in the AR transcript (Fig. 4D). Using primers that discriminate the two alternatively spliced isoforms, we found that the constitutive isoform connecting exon 2 and 3 decreased, whereas the alternative isoform connecting exon 2 and 2b increased upon HNRNPL knockdown (Fig. 4D). The pre-mRNA level of AR was not affected (Fig. 4D), confirming that HNRNPL regulates AR isoform expression through alternative splicing but not direct transcription. Moreover, RIP-qPCR analysis confirmed that HNRNPL binds to AR intron 3, contiguous with exon 3 that is alternatively spliced (Fig. 4 C and E). The regulation of the protein level of the constitutive AR isoform by HNRNPL was further confirmed by Western blot analysis (Fig. 4F). In addition to AR, we also validated another HNRNPL bound and regulated alternative splicing target MYH10 (Fig. S4 A–C and Dataset S3), where HNRNPL knockdown significantly reduced the exon 6-containing isoform but slightly increased the exon 6-skipping isoform.

Fig. S4.

Splicing validation of MYH10 and HNRNPL does not affect DHT/AR target genes. (A) Genome browser representation of HNRNPL binding over MYH10 loci as determined by RIP-seq. The peak region is indicated by a red rectangle. (B) RIP-qPCR confirmation of HNRNPL binding on MYH10 intron. Data are shown as mean ± SEM, n = 3. (C) RT-qPCR confirmation of HNRNPL-dependent alternative splicing of MYH10. Data are shown as mean ± SEM, n = 3. (D) Box plot showing that there are no change in the expression levels of DHT up- or down-regulated genes upon HNRNPL knockdown. The DHT up- and down-regulated gene lists were obtained from Wang et al. (62). (E) Venn diagram of HNRNPL RIP-seq peaks between vehicle and DHT conditions. LNCaP cells were cultured in charcoal/dextran treated FBS-containing medium for three days and then treated with either vehicle (1:1,000 ethanol) or 10 nM DHT for 4 h before being harvested for HNRNPL RIP-seq analysis. (F) Knockdown of HNRNPL does not affect classic target gene expression induced by androgen/AR signaling. Control or HNRNPL knockdown LNCaP cells were treated with either vehicle or 10 nM DHT for 4 h and RT-qPCR analysis was performed to evaluate the indicated androgen/AR target gene expression. Data are shown as mean ± SEM, n = 3.

HNRNPL Regulates Circular RNA Formation.

Circular RNA (circRNA), generated by nonsequential back splicing of pre-mRNA (Fig. 5A), has drawn increasing attention in the RNA field (32–34). Increasing length and depth of high-throughput sequencing allowed increasing number of circRNAs to be discovered and characterized. Rather than splicing noise, circRNA has been shown to be tightly regulated and functional (35). Recent advances in the understanding of circRNA biogenesis suggested splicing factors as regulators involved in circRNA formation (32, 36, 37).

Fig. 5.

HNRNPL is involved in CircRNA regulation. (A) Schematic depiction of circRNA formation via back splicing of linearly aligned exons. (B) Correlation of fold-change in abundance of HNRNPL-regulated circRNA (y axis) and its cognate parental mRNA (x axis). Note the scale difference between the two axes. (C) Association of HNRNPL binding status with circRNA expression regulation. The fold-enrichment represents the fraction for each category over all circRNA genes. (*P < 0.05). (D) RT-qPCR analysis of several HNRNPL-regulated circRNAs and their cognate pre-mRNAs. Validated circRNAs are selected as more than ∼twofold change upon HNRNPL knockdown and with HNRNPL binding in any region of parental genes from Dataset S4. Data are shown as mean ± SEM, n = 3, *P < 0.05. (E) Genome browser representation of HNRNPL binding over PPFIA2 loci as determined by RIP-seq. The back-spliced exons that generate two different circRNAs (circ-1 and circ-2) are denoted. The two flanking HNRNPL-binding peaks are highlighted. (F) RIP-qPCR validation of two HNRNPL-binding peaks over PPFIA2 transcript highlighted in E. Data are shown as mean ± SEM, n = 3. (G) Sanger sequencing traces for confirmation of back-spliced exon junctions of the two separate PPFIA2 circRNA. (H) Schematic description showing the design of GAPDH minigene assay. (I) RT-qPCR analysis of relative circRNA formation normalized by pre-mRNA levels for different GAPDH minigenes transfected into HEK293T cells. Data are shown as mean ± SEM, n = 3. (J) CircRNA formation of the GAPDH minigene is dependent on HNRNPL. RT-qPCR analysis was performed in G4 minigene-transfected samples upon control or HNRNPL knockdown. Data are shown as mean ± SEM, n = 3.

To determine whether HNRNPL is involved in circRNA regulation, we profiled the circRNA transcriptome in LNCaP cells by RNase R enrichment and RNA-seq profiling in either control or HNRNPL knockdown conditions, and determined differential expression of circRNAs (Dataset S4). Among 4,389 detected circRNAs in the LNCaP transcriptome, we identified 139 significantly up-regulated and 93 down-regulated circRNAs upon HNRNPL knockdown, whose parental mRNA gene expression remains unchanged (Fig. 5B). HNRNPL binding significantly associates with up-regulated circRNAs and to a lesser extent with down-regulated circRNAs upon HNRNPL knockdown when the binding events occur at both flanks or within the circRNAs (Fig. 5C). Using divergent PCR primers, we validated the expression change of several HNRNPL-regulated circRNAs upon HNRNPL knockdown, but their parental pre-mRNAs remain unchanged (Fig. 5D). The circularity and back splicing of these circRNAs were confirmed by RNase R treatment and Sanger sequencing (Fig. S5 A and B). Interestingly, one parental transcript can give rise to multiple circRNAs, as exemplified by PPFIA2 (Fig. 5 D–G). Between the two highlighted HNRNPL binding peaks in Fig. 5E, two HNRNPL-repressed circRNAs (denoted as circ-1 and circ-2) were present, indicating the in vivo complexity of circRNA regulation by HNRNPL.

Fig. S5.

CircRNA Regulation by HNRNPL. (A) RT-qPCR analysis of indicated circRNAs and mRNAs before and after RNase R treatment. (B) Sanger sequencing traces for confirmation of back-spliced exon junctions of indicated circRNAs. (C) Sanger sequencing traces for confirmation of back-spliced exon junction of the GAPDH circRNA from minigene assay. (D) Bar plot showing the relationship between the fractions of genes that form circular transcripts and the number of HNRNPL binding sites that they have.

To further test whether HNRNPL directly regulates circRNA formation via HNRNPL–RNA interaction, we designed a minigene system in which a fragment of GAPDH pre-mRNA with exon 5 and 6, as well as their flanking introns, was cloned into an expression plasmid. We then created different minigene constructs by introducing an HNRNPL binding site (CA)₂₀ to either or both of the flanking introns to evaluate the effect of HNRNPL binding on circRNA formation. These minigene constructs were transfected into HEK293T cells and relative circRNA levels were determined and normalized against pre-mRNA expression (Fig. 5H). This approach has several advantages in the evaluation of cis element or trans factor effects on circRNA formation, as there is no endogenous circRNA formation in the GAPDH locus and little endogenous pre-mRNA expression compared with the transfected minigene. HNRNPL binding on the flanking introns significantly enhances circRNA formation, especially when present at both sides (Fig. 5I), and HNRNPL knockdown significantly reduces the circRNA expression (Fig. 5J). Sanger sequencing confirmed the back-splicing exon junction of this GAPDH circRNA (Fig. S5C). Consistent with this, we found that one or more HNRNPL binding sites correlates with an increased likelihood to form circRNAs in our LNCaP cell profiling data (Fig. S5D). Because this may reflect the co-occurrence of HNRNPL with other key circRNA-promoting factors, integrating data on the other factors might be needed to better elucidate the role of HNRNPL in enhancing or repressing circRNA expression in vivo.

HNRNPL and Its RNA Targets are Clinically Relevant in Human Prostate Cancer.

To evaluate the clinical relevance of HNRNPL and its RNA targets, we analyzed the mRNA expression of HNRNPL in several independent prostate cancer cohorts (38–41). HNRNPL expression is consistently higher in prostate tumors versus normal prostate (Fig. 6A). Moreover, we confirmed its higher protein expression in cancer by immunohistochemical analysis of tissue microarrays with 79 pairs of matched prostate tumor and benign samples (Fig. 6 B and C and Fig. S6 A and B).

Fig. 6.

HNRNPL and its RNA targets are clinically relevant. (A) Violin plot showing the HNRNPL expression difference between normal prostate and prostate tumors using the gene-expression data from three independent cohorts. TCGA dataset were from prostate adenocarcinoma data in The Cancer Genome Atlas (TCGA). The Agell/Brase datasets were combined because of the limited sample number in each individual dataset. (B) Representative images of tissue microarray analysis at 20× magnification staining HNRNPL in both tumor and matched benign prostate tissues. (Scale bar, 200 μm.) (C) Statistics of tissue microarray analysis on HNRNPL protein expression. Seventy-nine pairs of tumor and matched normal biopsies were analyzed. Combined score adapted from the Allred scoring system taking into account both staining intensity and percentage was used to quantify the protein expression. (D) Gene-signature association network between HNRNPL-regulated RNA targets (either alternatively spliced genes from Dataset S3 or circRNA genes from Dataset S4) and published gene signatures deposited in Oncomine database. Each node represents a gene set signature and each edge represents a significant association between two connected gene sets (P ≤ 4.91E-4). The size of node is proportional to the gene numbers within gene sets. The thickness of the edge is proportional to the significance of the association with thicker edge meaning more significant association. (E) Representative gene heatmap showing the HNRNPL-regulated alternatively spliced genes or circRNA genes are correlated with overexpressed gene signatures in prostate cancer versus normal prostate in Vanaja Prostate cohorts in D.

Fig. S6.

Clinical relevance of HNRNPL. (A) Percentage of HNRNPL positive staining as determined by tissue microarray analysis. (B) Mean intensity of HNRNPL staining as determined by tissue microarray analysis. (C) Enrichment plot by gene set enrichment analysis. HNRNPL-regulated circRNA genes (Dataset S4), but not HNRNPL-regulated alternatively spliced genes (Dataset S3), are significantly enriched for LNCaP essential genes. (D) Genomic alterations of HNRNPL across several cancer types from cBioPortal.

Using Oncomine Concepts Map analysis (42), we found that both HNRNPL-regulated alternatively spliced genes and HNRNPL-regulated circRNA genes are significantly associated with overexpressed gene signatures in prostate cancer versus normal prostate in several independent cohorts (Fig. 6 D and E). Furthermore, HNRNPL-regulated circRNA genes are significantly enriched among the LNCaP essential gene set (Fig. S6C), indicating a functional link between HNRNPL and its downstream RNA targets. The association of both HNRNPL expression and its target RNA expression with prostate cancer progression suggests that HNRNPL may regulate prostate cancer growth through these downstream RNA targets. In addition to increased expression in prostate cancer, HNRNPL also exhibits frequent genomic amplification in many other tumor types (Fig. S6D), indicating that its importance may not be limited to prostate cancer.

Discussion

Posttranscriptional events, including alternative splicing and circular RNA formation, are increasingly appreciated to play important roles in cancer progression. Factors regulating RNA biogenesis including RBPs have been implicated in multiple fundamental biological processes, both physiologically and pathologically. However, which and how RBPs are functionally important in the context of prostate cancer still remains largely unknown. To this end, we systematically identified the genes that are essential for prostate cancer growth by a genome-wide CRISPR screen and found HNRNPL to be a top essential RBP.

To determine the landscape of HNRNPL-associated RNAs, we performed HNRNPL RIP-seq and found that HNRNPL binds preferentially to CA-repeats or CA-enriched RNA motifs in introns and 3′UTR regions. Although these observations agree with two recent studies of HNRNPL–RNA interactions by cross-linking immunoprecipitation (CLIP)-seq in T cells and iCLIP-seq in HeLa cells, respectively (43, 44), we identified different RNA targets. This result may be because of context-specific HNRNPL–RNA interactions, as our RIP-seq data are aimed to understand HNRNPL function in prostate cancer.

Several hnRNP family proteins are known to function in alternative splicing (45). HNRNPL has been shown previously to enhance and repress alternative splicing of different targets (30, 46). Here we systematically profiled HNRNPL-dependent alternative splicing events in prostate cancer cells and integrated this with RIP-seq data to obtain a comprehensive view of direct HNRNPL-regulated alternative splicing. The identification of AR as an HNRNPL-regulated splicing target is interesting, in light of the central role of AR in prostate cancer growth. However, it is unlikely that HNRNPL loss-of-function decreases cell growth uniquely through the regulation of AR, for the following reasons. First, overexpression of the full-length AR only partially rescues the cell-growth phenotype caused by HNRNPL knockdown (Fig. S3D). Second, the magnitude of full-length AR reduction is not sufficient to significantly affect androgen/AR-regulated target gene expression, nor does androgen affect genome-wide HNRNPL–RNA interactions (Fig. S4 D–F). Third, despite lower expression of the full-length AR upon HNRNPL knockdown, the increase of the short AR isoform, corresponding to a constitutively active variant that could bind to DNA and execute transcription without androgen stimulation, would be expected to compensate for the reduction of full-length AR in regulating gene expression and cell growth (47). Fourth, HNRNPL is required in AR-null prostate cancer cell lines (Fig. S2C), indicating that AR is not the only key target of HNRNPL. Consistently, the HNRNPL-regulated splicing targets are collectively associated with the overexpressed gene signatures in prostate cancer, implying that these targets may together contribute to mediate HNRNPL function (Fig. 6 D and E).

A very limited number of factors have been shown to regulate circRNA formation (32, 36, 37). Our study added HNRNPL to this list and suggested that a broader range of RBPs, especially splicing factors, might be involved in circRNA regulation. Minigene assays of HNRNPL confirmed the previous model that trans factor binding or direct complementary sequence pairing at two flanking introns could promote circRNA formation by bringing back spliced exon ends into close proximity (32, 36, 37). However, the pattern of circRNA regulation in vivo may prove to be more complicated because HNRNPL could either positively or negatively interfere with the splicing machinery that is responsible for circRNA generation. The diverse association pattern of HNRNPL binding and circRNA regulation also implies that trans factors do not have to bind only at the flanking introns to affect circRNA expression, broadening the way that circRNAs are regulated. It is worth noting that the HNRNPL-regulated circRNA genes are relevant to prostate cancer progression, implying the potential function of the circRNAs generated from these genes.

Considering the many other proposed functions of HNRNPL, such as affecting RNA half-life or translation efficiency through 3′UTR binding (28, 48), our investigation here may not be exhaustive. We found that HNRNPL regulates its RNA targets through either linearly alternative splicing or back-spliced circRNA formation, which may collectively contribute to its essentiality in prostate cancer. Our study not only identified a cohort of prostate cancer dependency genes, but also revealed how HNRNPL orchestrates RNA biology, suggesting that targeting RBPs or RBP–RNA interactions may be a novel therapeutic approach in prostate cancer.

Materials and Methods

Cell Culture.

LNCaP, CWR22Rv1, DU145, and PC3 cells were cultured in RPMI medium 1640 supplemented with 10% FBS. RWPE-1 cells were cultured in Keratinocyte Serum Free Medium (K-SFM; Kit Cat. no. 17005-042) with kit-supplied bovine pituitary extract and human recombinant epidermal growth factor. The use of human prostate samples for tissue microarray analysis has been approved by The Gelb Center Committee at Dana-Farber Cancer Institute.

CRISPR Screen.

For the pooled genome-wide CRISPR screen, 1 × 10⁸ LNCaP cells were infected with the pooled lentiviral GeCKO v2 library at a multiplicity of infection of 0.5. After 3 d of puromycin selection, half of the surviving cells were stored as 0-d control samples, and the rest of cells were cultured for an additional 2 wk. PCR was performed on genomic DNA to construct the sequencing library. Each library was sequenced at 30∼40 million reads to achieve ∼300× average coverage over the CRISPR library and data were analyzed by MAGeCK and MAGeCK-VISPR (21, 22).

siRNA Knockdown.

LNCaP cells were seeded in 24-well plates and transfected with 20 nM siRNA oligos using RNAiMax reagent (Life Technology). Knockdown efficiency was determined after 72 h of transfection. Cell counting was performed after 6 d of transfection. Detailed siRNA target sequences can be found in SI Materials and Methods.

RNA Isolation and qRT-PCR.

RNA was isolated using the RNeasy Mini Kit (Qiagen). Reverse-transcriptase (Invitrogen) was used for random-primed first-strand cDNA synthesis. Real-time PCR was carried out on ABI Prism 7300 detection system using SYBR Green PCR master mix. The ΔΔCt method was used to comparatively quantify the amount of mRNA level. RPS28 gene expression served as the internal control. Primer sequences can be found in SI Materials and Methods.

RIP.

Cells were cross-linked with 0.3% formaldehyde for 10 min at room temperature and lysed with RIPA lysis buffer for 10 min on ice before sonication. The supernatant was collected and incubated with RIP antibodies for 4–6 h. After washing with RIPA buffer, RNA was eluted from the beads with 100 µL NaHCO₃ and 1% SDS in the presence of proteinase K and RNase inhibitor at room temperature for 10 min with occasional vortex. The eluted material was decross-linked at 65 °C for 45–60 min before purification of RNA using TRIzol LS reagent (Life Technology). DNase I treatment was performed to remove any residual DNA. RIP RNA was used for either library preparation or direct qPCR assay.

Next-Generation Sequencing.

All RIP-seq and RNA-seq libraries were constructed with TruSeq Stranded mRNA Library Prep Kit from Illumina according to the manufacturer’s manual. RIP-seq libraries were sequenced as single end reads in duplicates, whereas RNA-seq libraries were sequenced as paired end reads in triplicates (splicing analysis) or duplicates (circRNA analysis) at ∼50 million reads per library using either HiSEq. 2000 or NextSEq. 500 sequencing platforms.

Bioinformatics Analysis and Statistics.

De novo motif analysis was performed using MDscan and SeqPos that were implemented in the Cistrome package. The RIP-seq and RNA-seq reads were aligned against the hg19 human reference genome with the University of California, Santa Cruz (UCSC) known gene transcript annotation using TopHat v2.0.9. Because the RIP-seq and RNA-seq data are strand-specific, the RIP-seq peaks were identified for “+” and “–” strands separately using macs2 v2.0.10, with a scanning window size of 100 bps but without shifting the reads [peak length ≥ 150, false-discovery rate (FDR) ≤ 0.01, fold-change ≥ 4, and RIP read count ≥ 50]. The differential alternative splicing analysis was performed using rMATS, a computational tool that allows for the analysis of replicate RNA-seq data. The circular RNA was identified using the CIRCexplorer method (see details in SI Materials and Methods). The differential expression analysis of circRNAs was performed using LIMMA (FDR ≤ 0.35, fold-change ≥ 2), and the total normalized count no less than 15. Moreover, only those circRNAs, whose parental genes showed little expression change (fold-change ≤ 1.2), were considered differentially expressed. Notably, increasing sequencing depth would further enhance the power and robustness to identify or calculate the differential expression of circRNAs. Statistical analysis of tissue microarray data were performed using SPSS software, v22.0 (SPSS).

SI Materials and Methods

Cell Culture, Reagents, and Antibodies.

LNCaP, CWR22Rv1, DU145, and PC3 cells were cultured in RPMI medium 1640 supplemented with 10% FBS. RWPE-1 cells were cultured in K-SFM (Kit Cat. no. 17005-042) with kit-supplied bovine pituitary extract and human recombinant epidermal growth factor. RNase R was purchased from Epicentre (Cat. no. RNR07250). The antibodies were purchased from the following sources: GAPDH (FL-335, Santa Cruz, Cat. no. sc-25778), HNRNPL (4D11, Santa Cruz, Cat. no. sc-32317; D5, Santa Cruz, Cat. no. sc-48391), HNRNPK (3C2, Santa Cruz, Cat. no. sc-32307), and AR (N-20, Santa Cruz, Cat. no. sc-816).

Pooled Genome-Wide CRISPR Screen.

LNCaP cells were cultured in RPMI medium 1640 supplemented with 10% FBS and infected with the pooled lentiviral library (GeCKO v2 library) at an multiplicity of infection of 0.5. Large-scale spin-infection of 1 × 10⁸ cells were carried out in four 12-well plates with 2 × 10⁶ cells per well. Wells were pooled together into larger flasks on the day after spin-infection. After 3 d of puromycin selection, half of the surviving cells were stored as 0-d control samples, and the rest of the cells were cultured in white RPMI medium 1640 supplemented with 10% charcoal-stripped FBS in the presence of 10 nM DHT for an additional 2 wk. PCR was performed on genomic DNA to construct sequencing libraries with each containing around 300 μg DNA each. Each library was sequenced at around 30∼40 million reads to achieve ∼300× average coverage over the CRISPR library. Data analysis was performed by MAGeCK and MAGeCK-VISPR.

Targeted siRNA Knockdown for Functional HNRNP Genes.

LNCaP cells were seeded in a 24-well plate and transfected with 20 nM siRNA oligos by RNAiMax reagent (Life Technology, Cat. no. 13778-150). Knockdown efficiency was determined after 72 h of transfection. Cell counting was performed after 6 d of transfection. The siRNA oligos targeting the HNRNP genes were purchased from Sigma and Dharmacon. The siRNA target sequence for siControl is 5′- GCGACCAACGCCUUGAUUG-3′. The siRNA target sequences for the HNRNP genes are as follows:

Gene name	siRNA-a	siRNA-b
HNRNPK	GCAAGAAUAUUAAGGCUCU	GAUCUUGGUGGACCUAUUA
HNRNPL	CAUCAUGCCUGGUCAGUCA	AGGUUUGUAGAGGCUUACU
HNRNPC	CAGUAGAGAUGAAGAAUGA	GAUGAAGAAUGAUAAGUCA
RBMX	CGAUAGAGAUGGAUAUGGU	CUACUCAAGUGGUCGUGAU
HNRNPDL	GUCACUAUGGAGGAUAUGA	CAAGGAUAUGGAAAUUACA
PCBP1	CGGUUAAGAGGAUCCGCGA	GUAUUAGUCUGGCCCAGUA
HNRNPM	GAUUGACGUUCGAAUUGAU	CGAUUUGGAUCUGGGAUGA
FUS	CAGAGCUCCCAAUCGUCUU	GGCUAUGGAACUCAGUCAA
PCBP2	GCAUUAGCCUGGCUCAAUA	GAACCCAGUGGAAGGAUCU
HNRNPU	GUGGAAUCGGCUAUCCAUA	GUCACUAACUACAAGUGGA
HNRNPUL1	CUAUAUCCUAGAUCAGACA	GUUGCUAUUGACACCUAUA
HNRNPF	GGUACAUUGGCAUCGUGAA	CAAUAUGCAGCACAGAUAU
SYNCRIP	GCUAGUUGCACAUAGUGAU	GUUAUGCGUUUGUCACUUU
HNRNPH3	GACAGUACGACUUCGUGGA	CAAUUACAGUGGAGGAUAU
PTBP1	CAAGAACUUCCAGAACAUA	CUGACCAAGGACUACGGCA
HNRNPH2	GUACAUUUGUGGGAGUUGA	CUGUACAUUUGUGGGAGUU
HNRNPA3	GUACAUUCCUGAGGUCUUU	CAAUGUGUGCUCGACCACA
HNRNPA0	CAGACCAAGCGCUCCCGUU	CACUUUGAGGCCUUUGGGA
RALY	GGCAAGCAAUGUAACCAAC	GCAAGCAAUGUAACCAACA
HNRNPH1	CUUCUUGAAUUCUACAGCA	CUUUGUACGGCUUAGAGGA
HNRNPD	GUAGUAGUGUCUUACUGGU	GUUGGGUCCCUCUGAAGUU
HNRNPUL2	CUUAGUAAGCUGGUCCAGA	CUAUAUGGAUGAGGUGACA
HNRNPA2B1	CUCAUGUAACUGUGAAGAA	CUGAUAGGCAGUCUGGAAA
HNRNPLL	AGUGCAACGUAUUGUUAUA	CUUAAUGUUUGCGUGUCUA
HNRNPA1	GGAAGAGUUGUGGAACCAA	GUGGUAAUGAGAGAUCCAA
HNRNPR	GGAGUAUGGAGUAUGCUGU	GCUAGUGCUUUGUCUUAGU
HNRNPAB	GAGAUUGAGGCCAUUGAAU	CUGUGGACCCUGGUUGUAA

RNA Isolation and qRT-PCR.

RNA was isolated using the RNeasy Mini Kit (Qiagen). Reverse transcriptase (Invitrogen) was used for random-primed first-strand cDNA (cDNA) synthesis. Real-time PCR was carried out on ABI Prism 7300 detection system using SYBR Green PCR master mix. The ΔΔCt method was used to comparatively quantify the amount of mRNA level. RPS28 gene expression served as the internal control. Primer sequences detecting mRNA levels are listed below:

Gene name	Forward primer (5′ to 3′)	Reverse primer (5′ to 3′)
RPS28	CGATCCATCATCCGCAATG	AGCCAAGCTCAGCGCAAC
HNRNPK	AGCTCCCGCTCGAATCTGAT	CCTCAACTCGCAGTCAAAGTC
HNRNPL	TTCTGCTTATATGGCAATGTGG	GACTGACCAGGCATGATGG
HNRNPC	GGAGATGTACGGGTCAGTAACA	CCCGAGCAATAGGAGGAGGA
RBMX	GCTCTTCATTGGTGGGCTTA	GGGCTTTCAAAGGTGACAAA
HNRNPDL	TGTGAGATCACCCGTTGTGT	CAGGTTTCAGAGGACCTGGA
PCBP1	AAAGGCGGGTGTAAGATCAAAG	GGCAAATCTGCTTGACACACTC
HNRNPM	TGGTCCGAGCAGACATTCTTG	TGACGTGCATTGGTCTATCAAA
FUS	TCAATCCTCCATGAGTAGTGGT	CACGGTCCTGCTGTCCATA
PCBP2	GCGCAGATCAAAATTGCGAAC	ATATTGAGCCAGGCTAATGCTG
HNRNPU	GAGCATCCTATGGTGTGTCAAA	TGACCAGCCAATACGAACTTC
HNRNPUL1	GAAGCACCTTCCGTCTACAGA	AGGAGAAAGGCTCTTCGCCTA
HNRNPF	CTGCTCTGTTGAGGACGTG	CCTGCCCTCTCTAGTGTAGATG
SYNCRIP	GAGCTAGAGGAAGGGGTGGT	CTCTTTGTTGTTGGGCACCT
HNRNPH3	AATGGTCCAAATGACGCTAGTG	CTCCCCTGGTAGTCCATCGT
PTBP1	AGCGCGTGAAGATCCTGTTC	CAGGGGTGAGTTGCCGTAG
HNRNPH2	GAAGCATACAGGTCCGAATAGC	CGCCCCTGAAAGTCCACTG
HNRNPA3	TGATGGGCGTGTAGTGGAAC	AGCAGACTGCATCTCTTGTTTAG
HNRNPA0	TGGCTTCGTGACCTACTCCAA	GGCCTCCGACAAAGAGCTT
RALY	TTCAGGCAAGCAATGTAACCA	CACGGCCATACTTAGAGAAGATG
HNRNPH1	ATTCAAAATGGGGCTCAAGGTAT	GTGTCAGGACTATTTGGACCAG
HNRNPD	GCGTGGGTTCTGCTTTATTACC	TTGCTGATATTGTTCCTTCGACA
HNRNPUL2	GGCAAAGGTAACCCAGAATCTC	GGACGGGAAAAATCAACAGACC
HNRNPA2B1	AGCTTTGAAACCACAGAAGAA	TTGATCTTTTGCTTGCAGGA
HNRNPLL	ACCATTCCTGGTACAGCACTG	TGGCCAGCACTTGTAAAGC
HNRNPA1	TCAGAGTCTCCTAAAGAGCCC	ACCTTGTGTGGCCTTGCAT
HNRNPR	GCAAGGTGCAAGAGTCCACA	CACGCCAGAGTACACACTGTC
HNRNPAB	ATTGAGGCCATTGAATTGCCA	GGCCACCTTGATCTCACACTT
KLK3	TGTGTGCTGGACGCTGGA	CACTGCCCCATGACGTGAT
TMPRSS2	GGACAGTGTGCACCTCAAAGAC	TCCCACGAGGAAGGTCCC
FKBP5	GCGGAGAGTGACGGAGTC	TGGGGCTTTCTTCATTGTTC

RIP.

Adherent cells grown in 15-cm plates were first cross-linked with 0.3% formaldehyde for 10 min at room temperature before the reaction was quenched by adding one-tenth volume of 1.25 M glycine for 5 min. Cells were then scraped off the plates and lysed with RIPA lysis buffer (50 mM Tris, pH7.6, 150 mM NaCl, 1 mM EDTA, 0.1% SDS, 1% Nonidet P-40, 0.5% sodium deoxycholate, protease inhibitor, and RNase inhibitor) for 10 min on ice. Then the lysate was sonicated to assist solubilization and RNA fragmentation before centrifugation at 20,000 × g for 10 min at 4 °C. The supernatant was collected and precleared with protein G beads. The input fraction was obtained from the supernatant after preclear step. The antibodies were preincubated with protein G beads and washed with RIP wash buffer (RIPA lysis buffer without inhibitors) before adding the precleared cell lysate. After 4- to 6-h incubation, the beads were washed twice with RIPA lysis buffer followed by three washes with 1 M RIPA buffer (1 M NaCl in RIPA lysis buffer). The RNA was eluted from the beads with 100 µL NaHCO₃ and 1% SDS in the presence of proteinase K and RNase inhibitor at room temperature for 10 min with occasional vortexing. The eluted material was decross-linked at 65 °C for 45–60 min before purification of RNA using TRIzol LS reagent (Life Technology). DNase I treatment was performed to remove any residual DNA before phenol/chloroform/ethanol purification of the final RNA. RIP RNA can be used for either library preparation or direct qPCR assay. HNRNPL (4D11) antibody was used for HNRNPL RIP-seq. The RIP-qPCR primers used in this study are as follows:

Gene name	Forward primer (5′ to 3′)	Reverse primer (5′ to 3′)
RPS28	CGTGGAATTCATGGACGAC	GCTTCTCGCTCTGACTCCAA
CTBP1	ACGTCTGTGCTGTGATGTCC	CGGATGTCATAGATGCCACA
ROR2	GTGTCATTCAATATTCTGTGTGTG	ACAGAGAACACACTTAGAGACACAA
STX3	GCCATGTTTTAGCTGTGTGG	TTGTTGCTGTTGGTTGTGGT
AR intron	CCCACCTTCTCCAGTCTGTC	GTTGCTGAGTCAGGGGAAAG
PPFIA2 site 1	GACAAAAGTAAGGTCCAAATGGT	CCCCACCACAATTCTGTTTT
PPFIA2 site 2	GAGCAACTGGAAAGGCAAAG	TATCCAGCCACAATCGAACA
MYH10 intron	ACCTGGTTGTATCCCCTGTG	TCCACCATCACAGATAAGCAA

RNA Pulldown Assay.

RNA fragments were in vitro transcribed from corresponding PCR products amplified from genomic DNA by T7/T3 RNA polymerase. T7/T3 promoter sequences were introduced into PCR products via primers to generate strand-specific RNA fragments. Biotin-labeled UTP was incorporated into final RNA fragments during transcription. Cells were lysed with RIPA lysis buffer (50 mM Tris, pH7.6, 150 mM NaCl, 1 mM EDTA, 0.1% SDS, 1% Nonidet P-40, 0.5% sodium deoxycholate, protease inhibitor and RNase inhibitor) and then incubated with indicated biotin-labeled RNA fragments for 1 h at room temperature followed by Streptavidin beads incubation for 1 more hour. Beads were washed twice with RIPA lysis buffer and three times with 500 mM NaCl RIPA buffer. The associated proteins were analyzed by Western blot analysis. The PCR primers used to amplify in vitro transcription templates are as follows:

• EGFP: 5′- GATCACAATTAACCCTCACTAAAGGGATGGTGAGCAAGGGCGAGGAGC-3′ and 5′- GATCACTAATACGACTCACTATAGGGTTATCTAGATCCGGTGGATCCC-3′;

• RPS28: 5′- GATCACAATTAACCCTCACTAAAGGGCCATCATGGACACCAGCCGTGTG-3′ and 5′- GATCACTAATACGACTCACTATAGGGAACTTGAAACACAAACGCTTTAT-3′;

• LARP: 5′- GATCACAATTAACCCTCACTAAAGGGCCTGGTGACTCGGACATTCCAGG-3′ and 5′- GATCACTAATACGACTCACTATAGGGTGATCCGCTGTGCGGCCACAGGTC-3′;

• CTBP1: 5′- GATCACAATTAACCCTCACTAAAGGGCAGATAACGTACACGGATGCCACAG-3′ and 5′- GATCACTAATACGACTCACTATAGGGGTGTGTGACATCTGTGCAGGCCCTG-3′;

• ROR2: 5′- GATCACAATTAACCCTCACTAAAGGGACCTTCTTACTGCCCCTTCTTCTTC-3′ and 5′- GATCACTAATACGACTCACTATAGGGTCTTTGTGTGTGTCTGAATATTCTG-3′;

• STX3: 5′- GATCACAATTAACCCTCACTAAAGGGTTGTAGGAATTGTGTCTGGAACC-3′ and 5′- GATCACTAATACGACTCACTATAGGGACAGCTCTCTGATATATCAAATTCC-3′.

Tissue Microarray Analysis.

The use of human prostate samples has been approved by The Gelb Center Committee. Immunohistochemical staining for HNRNPL was performed on 4-μm paraffin sections cut from two prostate tissue microarray (TMA) blocks provided by the Gelb Center Tissue Bank. The immunochemical stain was initially optimized on the bench, and then transitioned to the Leica BOND-III (Leica Biosystems) autostaining system. Immunostaining was performed on tissue sections following deparaffinization in two 5-min changes of xylene and rehydration through graded alcohols to distilled water. After blocking endogenous peroxidase activity, sections were subjected to heat-induced epitope retrieval in citrate buffer (pH 6.1) for 30 min. Following heat-induced epitope retrieval, the primary mouse monoclonal antibody targeting HNRNPL (4D11, sc-32317, Santa Cruz Biotechnology) was applied to the sections. Protein levels were examined using a dilution of 1:5,000 for 1 h at room temperature. Incubation with the biotinylated universal secondary antibody was then performed. Visualization was performed with 3,3′-diaminobenzidine (DAB) as the chromogen substrate. Once stained, both TMA slides were scanned on the Olympus BX-51 W1 microscope using Vectra 2.0.8 software (Perkin-Elmer). Cores that were disrupted or contained insufficient tissue were eliminated from the analysis. Finally, cores from 79 pairs of matched tumor and benign samples were valid for later analysis. Following the standard bright-field TMA scanning protocol, a chromogenic spectral library was composed using the spectra of both the counterstain (hematoxylin) and the immunostain (DAB). Tissue segmentation algorithms were subsequently constructed using inForm v2.0.2 (Perkin-Elmer). Initially, a training set comprising three classes of tissue was created (i.e., tumor, stroma, and other). Representative regions of interest for each of these classes were marked on 15–20 images from the TMAs. The software was trained on these areas and tested to determine how accurately it could differentiate between the tissue classes. Cell-segmentation algorithms were then constructed for nuclear and cytoplasmic populations. Cell-segmentation algorithms identified nuclei as pixels above the minimum signal value and cytoplasm as a two-pixel radius around each identified nucleus. The spectral library and algorithms were then run on all samples. Poorly segmented cores were manually corrected via touchscreen editing following pathology review. Statistical analysis was performed using SPSS software, v22.0 (SPSS). All reported P values were two sided. For all statistical analysis, P < 0.05 was considered significant.

RNA Interference.

In addition to the siRNA screen of HNRNPs, additional siRNA oligos were synthesized to silence indicated genes. These siRNA oligos were purchased from Dharmacon. The target sequences are as follows:

• siControl: 5′-GCGACCAACGCCUUGAUUG-3′;

• siHNRNPL-1 (siL-1): 5′-GAAUGGAGUUCAGGCGAUG-3′;

• siHNRNPL-2 (siL-2): 5′-CUACGAUGACCCGCACAAA-3′;

• siHNRNPK-1 (siK-1): 5′-GAUCUUGGUGGACCUAUUA-3′;

• siHNRNPK-2 (siK-2): 5′-GUCGGGAGCUUCGAUCAAA-3′.

Cell Growth and Soft Agar Assay.

For cell-growth assays, cells were plated in 24-well plates and transfected with indicated siRNA oligos in triplicate. Cells were counted after 6 d of transfection using a hemocytometer. Anchorage-independent cell growth in soft agar was performed in triplicate with 10,000 LNCaP cells per well suspended in 1.5 mL medium containing 0.35% agar spread on top of 1.5 mL of 0.7% solidified agar in six-well plates. Colonies were stained with Crystal violet and counted after 4 wk of plating.

Minigene Assay.

The GAPDH gene (NM_002046) fragment encompassing exon 5 and exon 6 along with their flanking introns was PCR amplified and cloned into pcDNA3.1⁺ plasmid using BamH I and EcoR I sites. The HNRNPL binding (CA)₂₀ sequence was introduced into either or both ends along with PCR primers used to amplify the GAPDH minigene. The qPCR primers used for the GAPDH minigene assay are as follows: circRNA of GAPDH, 5′-GCCAAAAGGGTCATCATCTC-3′ and 5′-TGGACTCCACGACGTACTCA-3′; pre-mRNA of GAPDH, 5′-GCGAGATCCCTCCAAAATCAA-3′ and 5′-CAGGGCTGAGTCAGCTTCCC-3′.

Alternative Splicing Analysis.

LNCaP cells transfected with either siCtrl or siHNRNPL(siL1+siL2) for 3 d were cultured in the presence of 10 nM DHT for 4 h before being harvested. Polyadenylated RNA was extracted and constructed into sequencing library for RNA-seq analysis. Triplicates were prepared for each condition. Computational data analysis is elaborated in Bioinformatics Analysis, below. For individual cases of validation, qRT-PCR analysis was performed.

The qPCR primers used for the splicing validation are as follows:

Gene name	Forward primer (5′ to 3′)	Reverse primer (5′ to 3′)
AR isoform-a	AAGACCTGCCTGATCTGTGG	CGAAGACGACAAGATGGACA
AR isoform-b	GTGGAAGCTGCAAGGTCTTC	GCCACACTCTAGAGCTGCAA
AR pre-mRNA	CTGTGGAGATGAAGCTTCTGG	GGGCCCTGAAAGGTTAGTGT
MYH10 isoform-a exon5/6	GTTTCACTGGTTTAGGCGAT	AGGTGCTGGGAAGACAGAAA
MYH10 isoform-a exon6/7	CACAGTCTTCGCATTTCCAA	ATCGCCTAAACCAGTGAAAC
MYH10 isoform-b	CACAGTCTTCGCATTTCCAA	AGGTGCTGGGAAGACAGAAA

Circular RNA Analysis.

LNCaP cells were transfected with either siCtrl or siHNRNPL(siL1+siL2) for 3 d before being harvested. Total RNA was extracted with TRIzol reagent (Life Technology). The Ribominus kit (Life Technology) was used to remove rRNA. RNase R treatment was performed with 20 U RNase R (Epicenter) per microgram RNA at 37 °C for 15 min. RNAs with or without RNase R treatment were purified and subjected to library preparation and next generation sequencing. Duplicates were performed for each condition. Computational data analysis is elaborated in Bioinformatics Analysis, below. For individual cases of validation, qRT-PCR analysis was performed. The qPCR primers used for the circRNA validation are as follows:

Gene name	Forward primer (5′ to 3′)	Reverse primer (5′ to 3′)
circ-PRKAR1B	TCCAGCTTCTCGAAGTGCTC	TTCCCAAGGACTACAAAACCA
circ-ZMIZ1	GATGGAGCTGGAGTGAGGTG	TCTGCAGAAGGACCAGGACT
circ-FOXJ3	GACAAGCCTGTCCATACAAACC	TCTGTCATTGAACAAATGTTTCC
circ-PPFIA2-1	TTCAGAGATTTCTTCTTCCTTTTC	AGGTAAGGGAGCGACTGAGG
circ-PPFIA2-2	CCCTGCAGGCATTTAATTCT	ATTGTTGCCTTGCGTGAAC
circ-CCNY	CCGAAGTGCCACCAGATTAT	AAATGGTGGAGCAGGAACTG
pre-PRKAR1B	ATGGGAGTCCGACTGTGAGT	CTCCTGAGTTCAAGCGATCC
pre-ZMIZ1	TCTCAGGCACACCTCATCTG	GGCCCTCAGTTGTTCTCAAA
pre-FOXJ3	GTTCCTGGAGTGCTCACACA	AGGCAAAATCTCCCTCCTTC
pre-PPFIA2	TCTTCCTTTTCTAGAAGTTGTTCC	TCCCCATTCTTTTCATTTTGA
pre-CCNY	AGGACGAGTGAGCAATGGAC	GGGCAGAGTCAGGTTGTCTG

Bioinformatics Analysis.

CRISPR screen data were analyzed by MAGeCK and MAGeCK-VISPR. De novo motif analysis was performed using MDscan (51) and SeqPos that were implemented in the Cistrome package (52). The RIP-seq and RNA-seq reads were aligned against the hg19 human reference genomes with the UCSC known gene transcript annotation using TopHat v2.0.9 (53, 54). Because the RIP-seq and RNA-seq data are strand-specific, the RIP-seq peaks were identified for “+” and “–” strands separately using macs2 v2.0.10 (55), with a scanning window size of 100 bps, but without shifting the reads (peak length ≥ 150, FDR ≤ 0.01, fold-change ≥ 4, and RIP read count ≥ 50). The differential alternative splicing analysis was performed using rMATS (31), a computational tool that allows for the analysis of replicate RNA-seq data. The circular RNA was identified using the CIRCexplorer (32). Briefly, RNA-seq reads were first mapped using TopHat 2.0.9 (parameters: -a 6–microexon-search -m 2) against the hg19 human reference genome with the UCSC known gene transcript annotation. Unmapped reads were then extracted and mapped onto the relevant reference genome using TopHat-Fusion (56) (TopHat 2.0.9, parameters:–fusion-search–keep-fasta-order–bowtie1–no-coverage-search). Back-spliced junction reads were extracted and further realigned against existing gene annotations to determine the precise positions of the donor or acceptor splice site for each back-spliced event. Junction reads with shifted alignments against canonical splice sites were adjusted to the correct positions and reads with alignments on different genes or noncanonical splice sites were discarded. Only those back-spliced junctions that were supported by at least one junction read in two replicates of RNase R-enriched RNA-seq were considered as candidate circRNAs. The back-spliced junction read count of individual circRNA across different samples was normalized by trimmed mean of M-values method implemented in edgeR (57, 58). The differential expression analysis of circRNAs was performed using LIMMA (59) (FDR ≤ 0.35, fold-change ≥ 2), and the total normalized count no less than 15. Moreover, only those circRNAs with parental gene showing little expression change (fold-change ≤ 1.2) were considered differentially expressed. Notably, increasing sequencing depth would further enhance the power and robustness to identify or calculate the differential expression of circRNAs. Genomic alterations of HNRNPL were analyzed using cBioPortal (60, 61).