A Genome-Wide CRISPR Activation Screen Identifies PRRX2 as a Regulator of Enzalutamide Resistance in Prostate Cancer

The androgen receptor (AR) is the single most important therapeutic target in prostate cancer. Hence, androgen-deprivation therapy (ADT) and potent AR pathway inhibitors (ARPI) such as enzalutamide are widely used to treat this disease (1). However, the majority of patients eventually develop resistance, at which point therapeutic options are scarce.

Most of the mechanisms underlying ARPI resistance rely on re-activation of the AR pathway. Therefore, the development of new and more potent anti-AR inhibitors is an active field of research (2). However, a growing number of patients develop alterations in pathways that bypass the need for AR activation, giving rise to androgen-independent (AI) ARPI-resistant neuroendocrine prostate cancer (NEPC) or double-negative prostate cancer (DNPC; refs. 3, 4). Furthermore, recent evidence suggests that enzalutamide treatment itself can select AR-low/indifferent preexisting clones, promote a lineage switch from luminal to basal phenotype or activate certain oncogenic pathways that contribute to disease progression (5, 6).

To identify genes that confer enzalutamide resistance, which may inform new therapeutic strategies, we used an in vitro genome-wide CRISPR activation (CRISPRa) screen in the androgen-sensitive LNCaP cell line. In addition, our cell line had an AR activity reporter [androgen-responsive element (ARE)-GFP], which allowed for classification of the population of enriched single-guide RNAs (sgRNA) between those with androgen-dependent (AD) versus AI mechanisms.

Our screen identified the paired-related homeobox2 (PRRX2) transcription factor (TF) as one of the top hits. PRRX2 is a target of the TGFβ pathway and helps mediate the epithelial-to-mesenchymal transition (EMT) in breast cancer (7). In this study, we show that PRRX2 is an oncogene in prostate cancer and that PRRX2 overexpression mediates enzalutamide resistance, which can be overcome via BCL2 and CDK4/6 inhibition.

LNCaP, 22RV1, and DU145 cells were obtained from the ATCC. LNRII cells were a generous gift from Dr. Jindan Yu at Northwestern University (Chicago, IL). All cells were grown in RPMI1640 media supplemented with 10% FBS and 1% penicillin/streptomycin antibiotic solution. For assays in charcoal-stripped serum (CSS) medium, cells were first hormone starved in RPMI1640 media without Phenol red supplemented with 10% CSS FBS and 1% penicillin/streptomycin for 72 hours before start of assay. LAPC9-AI cells were a generous gift from Dr. Dean Tang group at Roswell Park. These cells were maintained in castrated mice. All cells were verified as Mycoplasma-free and genetically authenticated by ATCC.

For the CRISPRa screen, we used the Calabrese library Set A which contains 3 different sgRNAs to activate each gene (8). Briefly, LNCaP cells were transduced with a GFP reporter under the control of an ARE (9). Then they were transduced with the dCas9-VP64-Blast and with the PXPR_502 vector for sgRNA expression at low multiplicity of infection (MOI). Cells were then treated with DMSO (Veh) or enzalutamide (25 μmol/L) during 7.5 or 9 weeks. Cells at 7.5 weeks were sorted into GFP(+) or GFP(–) using FACS. DNA was extracted, sequenced, and analyzed. For details, see Supplementary Materials and Methods.

To generate stable LNCaP-sgPRRX2 cells, we cloned the individual sgPRRX2 (Supplementary Table S1) into the PXPR_502 (Addgene #61425) backbone as described (8). For details, see Supplementary Materials and Methods.

To generate organoids, fresh tumors were shredded mechanically and then digested with collagenase (Gibco, 17018) in complete RPMI1640 media for 2 hours at 37°C. Then, they were incubated with trypsin for 5 minutes at 37°C. Digested tissues were then treated with DNase I (Sigma), and filtered with a 40-μm filter to obtain single cells. Cells were then frozen or used for organoid culture. For details, see Supplementary Materials and Methods.

NOD/SCID gamma (NSG) mice (Jackson Laboratory) used in this study were housed in a pathogen-free animal barrier facility. When mice were 6 to 8 weeks old, they were injected subcutaneously with different cell lines according to experiment. All experiments and procedures were performed in compliance with ethical regulations and the approval of the Northwestern University Institutional Animal Care and Use Committee. For details, see Supplementary Materials and Methods.

Enzalutamide (Axon Chemistry, 1613) 10 mg/kg or its vehicle (corn oil) was injected via intraperitoneal injection 3 times a week. Enzalutamide was dissolved in DMSO (2.5%) and corn oil.

Venetoclax (MedChem Express, HY-15531) was prepared in DMSO (5%), Tween 80 (5%), PEG300 (40%), and PBS (50%). 1 mol/L NaOH was added to adjust the pH to 7 and help solubilize the compound. Venetoclax 100 mg/kg or its vehicle was given via oral gavage 5 days on and 2 days off.

Palbociclib (MedChem Express, HY-50767A) was dissolved in 50 mmol/L sodium lactate pH 4, and was given via oral gavage 5 days on 2 days off. To avoid an interaction between them, venetoclax and palbociclib were administered with at least a 4-hour difference.

LNCaP cells were treated with DMSO control or enzalutamide 5 μmol/L for 7 days. Three biological replicates per condition were used for this experiment. For details, see Supplementary Materials and Methods.

For assay for transposase-accessible chromatin using sequencing (ATAC-seq) experiment, cells were treated with enzalutamide (5 μmol/L) or DMSO for 4 days. Two replicates of each condition were used for this experiment. For details, see Supplementary Materials and Methods.

For information regarding human prostate cancer datasets, see Supplementary Materials and Methods.

Gene enrichment analysis was performed using the Enrichr program (10–12). See Supplementary Materials and Methods for detail information.

For hierarchical clustering and similarity matrix construction, an input matrix with gene expression values of AR, NEPC, and PRRX2 signature genes was inputted into Clustergrammer (13).

Similarity matrix row distance was calculated using 1–cosine distance. For patient hierarchical clustering, the rows and columns were clustered using the cosine distance and average linkage parameters.

For more information, see Supplementary Materials and Methods.

To calculate the different scores, we downloaded the expression of the AR, PRRX2, and NEPC signature genes from cBioPortal. We then calculated the absolute value of the expression of each gene. Subsequently, we ranked the gene expression from 1 to n, where n = Number of patients. The highest rank was given to the patient with the highest expression of the gene. To obtain the final score, we added the ranks of the genes with a z-score >0 and subtracted the ones with z-score <0.

To classify androgen receptor–dependent prostate cancer (ARPC), NEPC, DNPC, and ARPC/NEPC populations from the Stand Up To Cancer (SU2C) 2019 in dataset from Fig. 5A, we normalized the scores from 0 to 1 and arbitrarily set up a threshold of 0.5. Patients with AR score and NEPC scores <0.5 were classified as ARPC (n = 85). Patients with NEPC score >0.5 and AR score <0.5 were classified within the NEPC group (n = 22). Patients with DNPC had both AR score and NEPC scores <0.5 (n = 100). Finally, patients with NEPC/AR had AR and NEPC scores >0.5 (n = 5).

Statistical analyses were performed using unpaired two-tailed Student t test or one-way ANOVA, unless otherwise indicated. Survival studies were analyzed by log-rank (Mantel–Cox) test. Correlations were analyzed by Spearman correlation coefficient (r). Data are presented as mean ± SEM, unless otherwise indicated. For all analyses, results were considered statistically significant with *, P < 0.05; **, P < 0.01; ***, P < 0.001; and ****, P < 0.0001.

Detailed protocols for these techniques are described in Supplementary Materials and Methods.

The data analyzed in this study were obtained from Gene Expression Omnibus (GEO). The RNA sequencing (RNA-seq) and ATAC-seq generated in this study are available in GEO with the accession number: GSE199539. For analysis using the Stelloo RNA-seq and chromatin immunoprecipitation sequencing (ChIP-seq), data were obtained from the gene ontology database with the accession numbers: GSE120741 and GSE120738, respectively (14). The Pomerantz ChIP-seq data was obtained from GSE70079 (15). ChIP-seq data from the PC3 and LNCaP cell lines were obtained from (GSM3145503; ref. 4) and (GSE14092; ref. 16) respectively. Patient data of paired patients before and after ADT treatment was obtained from GSE48403 (17). All the other data supporting the findings of this study are available within the article and its Supplementary Data and from the corresponding author upon reasonable request.

To identify new targets of enzalutamide resistance, we conducted a CRISPRa screen using the hormone-sensitive prostate cancer cell line LNCaP. The positive selection screen allows for enrichment of cells that have upregulated the expression of resistance-mediating genes (Fig. 1A). In addition, to discern between AD and AI resistance, our cells expressed GFP under the control of the ARE (Supplementary Fig. S1A). We validated the reporter system by stimulating and inhibiting the AR using DHT or enzalutamide, respectively (Supplementary Fig. S1B).

We then transduced the LNCaP-ARE-GFP cells with the dCas9-VP64 vector and subsequently with the sgRNA library at a low MOI to obtain only one sgRNA per cell (Supplementary Fig. S1C; see Materials and Methods). Four days postinfection, we divided the cells into 2 groups: Vehicle (Veh) and enzalutamide and treated them for 7.5 and 9 weeks to allow for sgRNA enrichment. To separate AD from AI mechanisms of resistance, we sorted the enzalutamide-treated population at 7.5 weeks into GFP(+) and GFP(–) fractions (Fig. 1A).

Analysis of the distribution of the different sgRNA populations showed that all the enzalutamide-treated groups had wider distributions with enriched and depleted sgRNAs indicating the selection pressure exerted by enzalutamide (Fig. 1B). Comparison between the unsorted enzalutamide-All population to untreated (T0) or Veh groups revealed AR as the most enriched sgRNA after enzalutamide treatment (Fig. 1C and D). This result validates our experimental approach, as AR overexpression is one of the main mechanisms of enzalutamide resistance (18, 19). In addition, our screen identified other genes previously linked to enzalutamide resistance such as WNT1, ID1, ID3, and CCND2 (Fig. 1C and D; refs. 5, 20, 21). Overall, 422 guides were enriched [fold change (FC) >3.5)] and 1,040 (FC <3.5) guides depleted post-enzalutamide compared with Veh (Supplementary Table S2).

To identify genes specifically enriched in AD or AI populations, we compared the sgRNA distribution in the enzalutamide-GFP(+) and enzalutamide-GFP(–) groups. FACS sorting revealed that 89% of the cells were GFP(–), suggesting that only a small population of cells activated the reporter in the presence of enzalutamide (Supplementary Fig. S1D). Surprisingly, AR was also enriched in the GFP(–) population (Supplementary Fig. S1E and S1F). This could be due to an incomplete separation of GFP(+) and GFP(–) cells during FACS. Alternatively, the ARE reporter may not have been a faithful indicator of functional AR activity in some cells. Nevertheless, direct comparison between the enzalutamide-GFP(+) and enzalutamide-GFP(–) sgRNA reads revealed that AR was two-fold more enriched in the enzalutamide-GFP(+) than the enzalutamide-GFP(–) population (Fig. 1E; Supplementary Fig. S1G). We analyzed the distribution of nuclear factor I B (NFIB), a TF that coregulates the expression of AR target genes (22). We found that NFIB was 3.14 times more enriched in the enzalutamide-GFP(+) population. In contrast, ID1 and ID3 were more enriched in the GFP(–) compared with the GFP(+) population. These genes mediate enzalutamide resistance via an AI mechanism (5).

To identify the main pathways related to enzalutamide resistance we performed a gene enrichment analysis using enriched sgRNAs in the enzalutamide-All versus Veh groups as input. We found an overrepresentation of the TGFβ, TP53, and WNT–β-catenin signaling pathways (Fig. 1F). Of note, all these pathways have been linked to enzalutamide resistance and prostate cancer aggressiveness in different studies (23–26).

Among the top hits, we identified PRRX2 a homeobox TF that mediates EMT and metastasis in breast and colon cancer (7, 27). PRRX2 was enriched in our enzalutamide-treated population and in the GFP(–) versus GFP(+) population (Fig. 1C–E). Of note, the role of PRRX2 in prostate cancer is unknown. To explore the novel resistance mechanisms potentially mediated by PRRX2, we focused on this gene for the rest of the study.

To validate the CRISPRa screen results, we generated a stable LNCaP overexpressing PRRX2 cell line using our CRISPRa system (LNCaP-sgPRRX2) and a control cell line (LNCaP-sgNC). PRRX2 overexpression enhanced growth in the presence of enzalutamide and in androgen-deprived CSS media conditions (Fig. 2A and B). In addition, PRRX2 overexpression increased colony formation in DMSO and in enzalutamide-treated conditions compared with control cells (Fig. 2C). Furthermore, LNCaP-sgPRRX2 cells grew better as spheroids in 3D culture conditions in CSS media and in the presence of enzalutamide compared with the control cells (Fig. 2D).

Next, we knocked-down PRRX2 by shRNA in the enzalutamide-resistant 22RV1 prostate cancer cells, which express the AR variant ARV7. Downregulation of PRRX2 significantly impaired cell proliferation and colony formation of 22RV1 cells (Fig. 2E and F; Supplementary Fig. S2A), and sensitized these cells to enzalutamide (Fig. 2F). Interestingly, PRRX2 knockdown decreased the levels of AR and ARV7 in 22RV1 cells, which might partially explain why they became sensitive to enzalutamide (Supplementary Fig. S2B). These results indicate that PRRX2 mediates resistance to androgen deprivation and to enzalutamide in vitro.

To extend these results in vivo, we injected LNCaP-sgPRRX2 or LNCaP-sgNC cells subcutaneously into intact non-castrated mice and allowed for tumors to be established. To model the effects of ADT and enzalutamide treatment, we castrated the mice after the tumors reached approximately 400 mm³ and then treated them with enzalutamide (Fig. 2G; Supplementary Fig. S2C). LNCaP-sgPRRX2 xenografts were able to regrow after castration and in the presence of enzalutamide in contrast to LNCaP-sgNC cells (Fig. 2H). Of note, 83.3% (10/12) of the injected sites developed a tumor (> 100 mm³ after 1 month of injection) in the LNCaP-sgPRRX2 group, compared with only 33.3% (4/12) in the control group, indicating that PRRX2 is oncogenic in prostate cancer (Fig. 2I). Of note, LNCaP-sgNC tumors did not express PRRX2, suggesting that tumor growth on these cells was mediated by a PRRX2-independent mechanism (Supplementary Fig. S2D). Overall, these results show that overexpression of PRRX2 is able to drive resistance to enzalutamide in vitro and in vivo.

To gain further insight into the role of PRRX2 in enzalutamide resistance, we performed RNA-seq comparing LNCaP-sgPRRX2 versus LNCap-sgNC cells under DMSO (NC-D vs. PR-D) and enzalutamide-treated (NC-E vs. PR-E) conditions (Fig. 3A; Supplementary Table S3). We clustered all the genes into four different clusters (Fig. 3B). Cluster 1 (C1) contains genes that are repressed by enzalutamide and induced by PRRX2. In this cluster, we found that PRRX2 attenuates the repressive effect of enzalutamide. Of note, MYC target genes, mTORC1 signaling and androgen response genes are enriched in this cluster (Fig. 3C). Cluster 2 (C2) contains genes induced by PRRX2 regardless of enzalutamide treatment. Interestingly, genes involved in regulating cell-cycle pathways such as the E2F target genes and G₂–M checkpoint are enriched in C1 and C2, suggesting that PRRX2 plays an important role in the regulation of this process (Fig. 3C and D). Cluster 3 contains genes repressed by PRRX2. The low number of genes in this cluster indicates that PRRX2 is mainly a transcriptional activator. Interestingly, P16 (CDKN2A), a tumor suppressor gene that impairs cell-cycle progression, belongs to this cluster. Finally, in Cluster 4 (C4) we observed that PRRX2 cooperates with enzalutamide to further induce the expression of certain genes. Genes and pathways implicated in cancer progression and enzalutamide resistance, e.g., CXCR4, CD44, WNT, TGFβ, and EMT are enriched in this cluster (Fig. 3E).

To study the RNA-seq data in more detail, we compared the PR-D versus NC-D and PR-E versus NC-E cells using gene set enrichment analysis (GSEA). We found that EMT was the most significantly enriched pathway in both DMSO and enzalutamide conditions with PRRX2 overexpression confirming the role of PRRX2 in EMT. Interestingly, E2F target genes and G₂–M pathways were enriched only when comparing PR-E versus NC-E cells. In addition, pathways related to the immune system were also enriched in PRRX2 overexpressing cells (Supplementary Fig. S3A and S3B).

Of note, we found that the androgen response pathway was negatively enriched (NES = –1.34) in the PR-D versus NC-D cells (Fig. 3F), raising the possibility that PRRX2 promotes cell transition from an AD luminal to an AI basal phenotype. Supporting this hypothesis, PR-E cells had a negative enrichment score for the luminal phenotype (NES = –1.55) and a positive NES score for the basal phenotype (NES = 1.87) when compared with NC-E cells (Fig. 3G). To further explore the concept of an AI phenotype triggered by PRRX2, we used two gene signatures. One consists of upregulated genes in AI (ADT-resistant) versus AD (therapy-naïve), and the other one consists of genes upregulated in AI (castrated) versus AD (intact) xenografts (28). We found that in PR-E cells were enriched for both of these signatures compared with NC-E cells (Fig. 3G).

Together, these results show that in the presence of enzalutamide, PRRX2 helps maintain the activation of cell-cycle genes and further enhances the expression of pro-tumorigenic enzalutamide-induced genes/pathways.

Due to the importance of AR in prostate cancer and enzalutamide resistance, we next investigated the relationship between AR and PRRX2. We first analyzed the expression correlation between AR and PRRX2 across four different datasets and found a significant negative correlation between these genes (Supplementary Fig. S4A). Furthermore, AR activity [measured by an AR score (29, 30)] and PRRX2 expression were also negatively correlated in the The Cancer Genome Atlas (TCGA; Spearman r = –0.46) and the SU2C datasets (Spearman r = –0.34; Fig. 4A).

To analyze the link between AR activity and PRRX2 further, we studied AR chromatin occupancy using an AR ChIP-seq patient dataset from Stelloo and colleagues (14). We clustered these patients according to AR occupancy, measured by the number of AR peaks (ChIP-seq), and PRRX2 expression (RNA-seq) in each patient. As noted earlier (14), AR expression did not always correlate with AR chromatin occupancy (Fig. 4B), suggesting, that AR expression may not be an accurate measure of AR activity. Our clustering revealed two subgroups, C1 contains higher numbers of AR chromatin binding peaks and lower PRRX2 expression. In contrast, C2 has patients with lower number of AR peaks and higher PRRX2 expression (Fig. 4B), suggesting that patients with higher AR activity have lower levels of PRRX2 expression and vice versa.

On the basis of these results, we hypothesized that AR negatively regulates PRRX2. To address this hypothesis, we first examined changes in PRRX2 expression using an RNA-seq data from paired patient samples before and after ADT (17). We found an increase in PRRX2 expression after ADT treatment suggesting that PRRX2 expression is a potential AR target gene (Fig. 4C). Next, we analyzed AR binding into the PRRX2 promoter region using ChIP-seq patient and cell lines datasets. Of note, we found an AR peak near the PRRX2 promoter present in LNCaP cells. Importantly, the same AR peak was present in 11 of 14 patient samples from the Pomerantz dataset (Fig. 4D; ref. 15). We confirmed the presence of this peak using ChIP-qPCR in LNCaP cells (Supplementary Fig. S4B). To analyze the functional relevance of this peak, we knocked down AR/ARV7 in 22RV1 cells using siRNA and found an increase in PRRX2 mRNA and protein levels (Supplementary Fig. S4C and S4D). We also treated LNCaP cells with enzalutamide. However, we found no effect in PRRX2 expression levels suggesting that AR alone is not sufficient to regulate PRRX2 in this model (Supplementary Fig. S4E).

The TGFβ pathway is known to induce PRRX2 expression. SMAD3 is one of the main signal transducers of the TGFβ pathway (31). To study the relationship between TGFβ and PRRX2 in prostate cancer, we analyzed a ChIP-seq dataset of SMAD3 binding in PC3 cells (4).Of note, we found a SMAD3 peak near the PRRX2 promoter coinciding with the location of the AR peak, suggesting that they could potentially coregulate PRRX2 expression (Fig. 4D). This peak was confirmed in DU145 cells stimulated with TGFβ1 using ChIP-qPCR in DU145 cells (Supplementary Fig. S4B). Further analysis of prostate cancer datasets showed a strong correlation between the expression of PRRX2, the TGFB1 receptor 2 (TGFBR2; r = 0.5478) and of its ligand (TGFB1; r = 0.6260; Fig. 4E). This result was also validated in the Stelloo dataset where we observed that patients with the highest expression of PRRX2 had lower AR expression and higher TGFB1 levels (Fig. 4F). In addition, stimulation of DU145 cells with TGFβ1 lead to an increase in PRRX2 transcription (Supplementary Fig. S4F).

LNCaP cells are insensitive to TGFβ1 stimulation due to the lack of TGFBR2 expression, which may explain the lack of PRRX2 induction upon AR inhibition in LNCaP cells. To test if AR and the TGFβ pathway cross-regulate PRRX2, we used LNRII cells, a cell line derived from LNCaP that exogenously expresses the TGFBR2 gene, making it competent for TGFβ1 stimulation (32). We found that stimulation of LNRII cells with TGFβ1 or treatment with enzalutamide alone increased the levels of PRRX2 expression at mRNA and protein levels in LNRII cells. However, treatment with both enzalutamide and TGFβ1 led to the highest increase in PRRX2 mRNA and protein levels (Fig. 4G–H). This increase was accompanied by an increase in p-SMAD3 levels. Of note, triple immunofluorescence staining of AR, PRRX2, and p-SMAD2/3 proteins in LNRII cells revealed that cells with the lowest expression of AR expressed higher levels of both PRRX2 and p-SMAD2/3 (Supplementary Fig. S4G). Taken together, these data suggest that full PRRX2 induction in tumors requires AR inhibition and activation of the TGFβ pathway.

Given that AR negatively regulates PRRX2, we hypothesized that PRRX2 is mainly expressed in low AR settings. NEPC and DNPC are two subtypes of prostate cancer characterized by low AR expression or activity. To investigate if PRRX2 was overexpressed in these groups, we classified patients from the SU2C dataset into four categories (ARPC, NEPC, AR/NEPC, DNPC) according to their AR and NEPC scores (Fig. 5A). Analysis of PRRX2 expression in each of the clusters revealed that DNPC had the highest expression of PRRX2 (Fig. 5B). In addition, the frequency of patients with PRRX2 upregulation was highest in DNPC (24.4%) compared with NEPC (13.6%) and ARPC (9.7%; Supplementary Fig. S5A). Of note, LNCaP-sgPRRX2 cells treated with enzalutamide had a positive enrichment of a DNPC patient gene signature (Supplementary Fig. S5B).

To further study the clinical role of PRRX2 in prostate cancer, we used the 21 significantly upregulated genes (top 15th percentile of the upregulated genes) in the PR-D versus NC-D cells to create a PRRX2 signature (Fig. 5C). We then analyzed if the PRRX2 signature genes were coexpressed in patients with prostate cancer by performed a similarity matrix with the PRRX2, AR, and NEPC signature genes in four different prostate cancer datasets. The similarity matrix revealed three unique clusters corresponding to PRRX2, NEPC, and AR in all datasets (Fig. 5D; Supplementary Fig. S5C and S5D).

We then classified the patients according to the expression of the PRRX2 signature genes using unsupervised hierarchical clustering. Analysis of the TCGA and the Memorial Sloan Kettering Cancer Center (MSKCC), datasets revealed only two groups denominated PRRX2-high and PRRX2-low clusters (Supplementary Fig. S5E). Of note, the PRRX2-high clusters had lower expression of AR target genes. The SU2C, Beltran, and DKFZ datasets formed three clusters denominated ARPC, NEPC, and PRRX2 (Fig. 5E; Supplementary Fig. S5F). Similar to our previous results, patients within the PRRX2 cluster had lower expression of AR and NEPC target genes, further supporting our hypothesis about higher PRRX2 expression in patients with DNPC. In addition, we found a 71.7% overlap between the patients in the PRRX2 cluster and the patients within the DNPC cluster in the SU2C dataset (Supplementary Fig. S5G).

Further characterization of the PRRX2 cluster revealed an overrepresentation of bone tumor metastases (Supplementary Fig. S5H). Furthermore, PRRX2 expression itself was higher in bone samples compared with other tissues (Supplementary Fig. S5I and S5J). To further characterize the different clusters, we analyzed the expression of genes significantly overexpressed in the PRRX2 cluster compared with the ARPC and NEPC clusters. We found an enrichment of immune related pathway such as the IFNγ and alpha signaling in the PRRX2 cluster (Supplementary Fig. S5K). These results correlate with previous observations, which show that DNPC is associated with immune-related pathways (33). Overall, these results show that a PRRX2 signature is able to divide patients with prostate cancer within a specific subgroup that shares certain characteristics with the patients with DNPC, such as bone metastasis and enrichment of immune-related pathways.

To study the significance of PRRX2 score in prostate cancer prognosis, we calculated a PRRX2 score using the genes within the PRRX2 signature (Fig. 5C). Overall, our score was distinct from the AR and NEPC scores and matched with the patient stratification previously made by hierarchical clustering (Supplementary Fig. S5L–S5N). We then calculated the disease-free survival (DFS) between patients with the highest PRRX2 score (25th percentile) compared with the rest of the patients in the TCGA and the MSKCC datasets, but found no difference in DFS in any of these datasets (Supplementary Fig. S5O and S5P). Of note, patients with high clinical Gleason scores were mainly located within the PRRX2-high score group. This result was consistent in three different datasets (Supplementary Fig. S5Q–S5S). Furthermore, the gene expression of PRRX2 was higher in samples with Gleason score higher than 8 (Supplementary Fig. S5T).

Analysis of the SU2C dataset, which contains patients with advanced metastatic prostate cancer, showed no difference in overall survival (OS) between high (top 25th percentile) and low (rest of the patients) PRRX2 score patients (Supplementary Fig. S5U). However, stratification of the patients with DNPC using our PRRX2 signature score revealed that the OS of the patients with the highest PRRX2 score (25th percentile) was significantly lower compared with the rest of the patients (12.88 vs. 25.66 months; P = 0.0054; Fig. 5F).

To further study the role of PRRX2 in AR-negative cells, we knocked-down PRRX2 in DU145 cells, considered a model of DNPC (34). We found a significant decrease in colony formation in DU145-shPRRX2 compared with control cells using two different shRNAs (Fig. 5G). To gain more insight into the role of PRRX2 in DNPC, we used the LAPC9-AI, an enzalutamide-resistant DNPC patient-derived xenograft model (35). LAPC9-AI was derived from the AD parental cells, LAPC9-AD. LAPC9-AI xenografts grow in castrated mice and express low levels of AR target genes KLK3, FKBP5 and of the NEPC gene, SYP. Of note, we found that LAPC9-AI xenografts express significantly higher levels of PRRX2 compared with LAPC9-AD cells (Supplementary Fig. S5V).

Similar to our observations in DU145 cells, PRRX2 downregulation in LAPC9-AI cells led to a significant decrease in the organoid size compared with control cells (Fig. 5H; Supplementary Fig. S5W). We validated these results in vivo after injection of LAPC9-AI-shPRRX2–1 cells into castrated NSG mice. We observed that PRRX2 downregulation delayed tumor growth and decreased proliferation, as evidenced by lower Ki67 expression (Supplementary Fig. S5X), suggesting that PRRX2 is necessary for cell survival in AR negative/DNPC cells (Fig. 5I). Together, these results show that PRRX2 is an oncogene and might play a role in the aggressiveness of prostate cancer within the DNPC population.

We sought to study the transcriptional function of PRRX2 in more detail by ChIP-seq experiments but were limited by the lack of suitable ChIP antibodies. Instead, we conducted an ATAC-Seq experiment using the LNCaP-sgNC versus LNCaP-sgPRRX2 cells under enzalutamide treatment. We then annotated the genes that had ATAC-seq peaks and found 2069 (P < 0.05) differentially opened chromatin regions (Supplementary Table S4). We found that open chromatin regions in PRRX2 overexpressing cells corresponded to genes such as the WNT receptor, FZD1, and the TF TRSP1, which plays a role in cell proliferation. Closed chromatin regions corresponded to the proapoptotic cadherin (CDH11; Fig. 6A; ref. 36).

To gain more insight into the molecular mechanisms driven by PRRX2, we integrated the RNA-seq and the ATAC-seq data. We found an overlap of 333 differentially expressed genes between ATAC-seq (P < 0.05) and RNA-seq (P < 0.05; Fig. 6B). Pathway analysis using all of the 333 overlap genes revealed that the E2F TF network was within the top pathways, consistent with our RNA-seq analysis (Fig. 6C). In addition, analysis of the pathways enriched using just the upregulated genes (RNA-seq) and the open chromatin regions (ATAC-seq; n = 100 genes) revealed an enrichment of apoptosis block by BCL2 family proteins and apoptosis evasion in cancer pathways (Fig. 6D).

To validate these results, we analyzed the expression of RB1, which is upstream E2F, in LNCaP-sgPRRX2 and sgNC cells after enzalutamide treatment. RB1 loss via genetic mechanisms or via RB1 protein inactivation through hyperphosphorylation (hp-RB1) leads to the activation of E2F TFs and subsequent cell-cycle progression (37). We found that enzalutamide treatment decreased p-RB1 (inactive) levels in both cell lines. However, p-RB1 reduction was more pronounced in the control cells compared with the LNCaP-sgPRRX2 cells (Fig. 6E). In addition, LNCaP-sgPRRX2 cells had higher levels of hp-RB1 in androgen-deprived CSS media compared with control cells (Fig. 6F). Of note, P16 (CDKN2A), a tumor suppressor that inhibits CDK4/6-mediated RB1 phosphorylation, was significantly lower in the PRRX2-overexpressing cells compared with control (Fig. 6E) confirming that the LNCaP-sgPRRX2 cells had a more active cell cycle than control cells (Fig. 6G). Importantly, LNCaP-sgPRRX2 cells were more sensitive to the CDK4/6 inhibitor palbociclib than control cells (Fig. 6H), validating the functional significance of this pathway in PRRX2 expressing cells (Fig. 6G).

Genetic loss of RB1 and TP53 co-occur in a subset of patients with prostate cancer and is associated with resistance to enzalutamide (6). We hypothesized that PRRX2 expression leading to RB1 inactivation via hyperphosphorylation will phenocopy RB1 loss and cooperate with TP53 loss to promote enzalutamide resistance. To test this hypothesis, we knocked-down TP53 in LNCaP-sgPRRX2 cells (Supplementary Fig. S6A). Importantly, we found that LNCaP-sgPRRX2-shTP53 cells were the most resistant to enzalutamide compared with control cells and to each alteration alone (Fig. 6I; Supplementary Fig. S6B). Of note, the combination of enzalutamide and palbociclib, synergized to inhibit LNCaP-sgPRRX2/shTP53 cells (Fig. 6J).

In addition to the E2F pathway, we noted that antiapoptotic pathways were enriched in enzalutamide-treated PRRX2 overexpressing cells. Treatment with enzalutamide increased BCL2 levels in both cell lines. However, the increase was higher in the LNCaP-sgPRRX2 cells compared with control. Similarly, we found higher levels of BCL2 in LNCaP-sgPRRX2 cells grown in CSS conditions (Fig. 6K). Notably, enzalutamide and BCL2 inhibition using venetoclax cooperated in inhibiting LNCaP-sgPRRX2 organoids (Fig. 6L and M). In addition, BCL2 knockdown using shRNAs decreased cell proliferation of LNCaP-sgPRRX2 cells under enzalutamide treatment (Supplementary Fig. S6C and S6D).

To extend the translational relevance of these findings, we then studied the relationship between PRRX2 and the expression of E2F downstream genes (CCND2, CCND3) and of BCL2 in the SU2C dataset. We found a positive correlation between the expression of PRRX2 with CCND2 and with BCL2 (Supplementary Fig. S6E and S6F). Of note, CCND2, CCND3, and BCL2 were also overexpressed in the PRRX2 patient cluster (Supplementary Fig. S6G and S6H). Importantly, BCL2 protein levels were increased in the PRRX2-high cluster within the TCGA dataset (Supplementary Fig. S6I). Altogether, these results suggest that PRRX2 plays a role in the regulation of the cell cycle, via the RB1/E2F pathway and of cell survival via the BCL2 pathway.

To test whether CDK4/6 and BCL2 inhibition will reverse PRRX2-driven enzalutamide resistance, we established LNCaP-sgPRRX2 xenografts in NSG mice and castrated the mice when the tumor reached approximately 300 mm³. Following recurrence of tumors to ∼400 mm³ after castration, we started treatment with enzalutamide alone or combinations of enzalutamide with palbociclib, venetoclax, or both. We did not include a venetoclax or a palbociclib only group based on our in vitro results, which show that treatment with enzalutamide is necessary to induce E2F pathway activation and BCL2 expression.

In contrast to parental LNCaP cells that are sensitive to castration and enzalutamide treatment, we found that castration and enzalutamide were not able to slow down the growth of LNCaP-sgPRRX2 tumors (Fig. 7A). The addition of venetoclax and of palbociclib to enzalutamide improved the response by significantly delaying tumor growth. However, the triple combination had the highest effect in suppressing tumor growth (Fig. 7A). Importantly, the drug regimen administered was well tolerated by mice (Supplementary Fig. S7A).

Cell proliferation measured by Ki67 staining confirmed that enzalutamide treatment did not have any effect in tumor cell proliferation compared with the Veh group, consistent with the results in tumor growth (Fig. 7B and C). In contrast, enzalutamide + palbociclib and the triple combination had the highest effect decreasing cell proliferation. In addition, groups that received venetoclax treatment showed increased apoptosis (Fig. 7B and C). Overall, these results show that inhibition of CDK4/6 and BCL2 can overcome enzalutamide resistance in a prostate cancer model with high PRRX2 expression.

Despite good initial response to ARPI such as enzalutamide, resistance remains a major challenge for the treatment of CRPC. With the advent of new and more potent ARPI, resistance mediated via AI mechanisms is becoming more common, in particular through the development of AR-negative/low prostate cancer such as NEPC and DNPC. Understanding the molecular mechanisms driving enzalutamide resistance is important for the development of new therapeutic strategies to treat these patients.

This study sought to identify mediators of enzalutamide resistance via a genome-wide CRISPRa screen. Of note, the top hit of our screen was AR, reiterating the importance of AR as a therapeutic target even in late stages of prostate cancer. Other interesting hits were WNT1, ID1, and ID3. The WNT pathway has been previously linked to AI prostate cancer and to enzalutamide resistance, ID1 and ID3 proteins are TFs that are upregulated in enzalutamide-resistant prostate cancer models and confer resistance in vitro and in vivo (5, 20, 21).

In this study, we focused on the homeobox TF PRRX2. This gene is upregulated in human prostate cancer compared with healthy glands (38). We found that PRRX2 is mainly expressed in DNPC and a PRRX2 signature was able to stratify patients with DNPC with worse OS, suggesting that this gene is particularly important for AR-low tumors. PRRX2 can mediate EMT, suggesting that it could contribute to maintain the cells in an undifferentiated stem-like phenotype. Interestingly, PRRX2 is overexpressed in a cluster of prostate basal stem cells (39). In our study, PRRX2 overexpression decreased the expression of an AR target gene/luminal signatures. Furthermore, enzalutamide treatment accentuated the decrease in expression of luminal signature genes while increasing expression of basal signature genes. However, it does not appear that PRRX2 is sufficient to drive a lineage switch from AR positive to AR-low phenotype.

The higher expression of PRRX2 in AR-low/DNPC tumors could be explained by coregulation of PRRX2 by AR and the TGFβ pathway. Of note, PRRX2 is induced by TGFβ in several models (7). Interestingly, the TGFβ pathway is upregulated after enzalutamide treatment and has also been shown to mediate enzalutamide resistance (24, 26, 32). Of note, there is an ongoing clinical trial with the combination of the TGFβ receptor inhibitor, galunisertib, with enzalutamide for patients with mCRPC (NCT02452008).

Previous studies have linked the oncogenic effects of PRRX2 to activation of the WNT–β-catenin and EMT pathways (27, 40). In this study, we expanded the oncogenic role of PRRX2 to the regulation of the cell cycle and antiapoptotic BCL2-mediated pathways. Interestingly, other groups have shown that BCL2 is upregulated after enzalutamide treatment and is overexpressed in AI prostate cancer (35, 41). Here, we show that PRRX2 enhances the effect of enzalutamide-mediated BCL2 upregulation. Of note, the combination of venetoclax and enzalutamide is currently under clinical trial for patients with mCRPC (NCT03751436). Our study suggests that treatment with palbociclib and venetoclax will reverse enzalutamide resistance particularly in DNPC with high PRRX2 expression. Notably, the combination of palbociclib, venetoclax, and the anti-hormonal agent fulvestrant is currently undergoing a clinical trial for ER⁺ breast cancer (NCT03900884; ref. 42).

Y. Rodríguez reports grants from NCI, Prostate Cancer Foundation Challenge Award and grants from American Association of University Women (AAUW)–International Fellowship during the conduct of the study. M.I. Truica reports grants from NCI during the conduct of the study. V. Sagar reports grants from NCI and grants from Prostate Cancer Foundation Challenge Award outside the submitted work. B. Lysy reports grants from NCI and grants from Prostate Cancer Foundation Challenge Award during the conduct of the study. M. Hussain reports other support from Arvinas, AstraZeneca, Bayer, Genentech, Astellas, RTP, MLI Peer View, AstraZeneca, ReachMD, PrecisCa, Merck, Medscape, Tempus, Janssen, Pfizer, and other support from Novartis outside the submitted work; in addition, M. Hussain has a patent for UM-14437/US-1/PRO issued to IMBIO and a patent for PCT/US2011/053233 issued. H. Han reports grants from NIH and grants from Prostate Cancer Foundation Challenge Award during the conduct of the study. S.A. Abdulkadir reports grants from NCI and grants from PCF during the conduct of the study. No disclosures were reported by the other authors.

Y. Rodríguez: Conceptualization, formal analysis, investigation, methodology, writing–original draft, writing–review and editing. K. Unno: Investigation, methodology, writing–review and editing. M.I. Truica: Investigation, writing–original draft, writing–review and editing. Z.R. Chalmers: Data curation, software, methodology, writing–review and editing. Y.A. Yoo: Methodology. R. Vatapalli: Investigation, methodology. V. Sagar: Methodology. J. Yu: Investigation. B. Lysy: Methodology. M. Hussain: Investigation. H. Han: Investigation. S.A. Abdulkadir: Conceptualization, resources, formal analysis, supervision, funding acquisition, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing.

This work was supported by NCI grants P50CA180995, F30CA50248, and F30CA50196 and Prostate Cancer Foundation Challenge Award. The authors thank Dr. Matt Clutter from the Northwestern High-throughput Core facility for helping with the CRISPRa screen design and execution; Dr. Ching Man Wai, Dr. Matthew J. Schipma, and Priyam Patel at the NUSeq Facility for ATAC-seq execution and data analysis; Carolina Ostiguin from the The Robert H. Lurie Comprehensive Cancer Center Flow Cytometry Core for help in flow cytometry analysis; The American Association of University Women (AAUW)–International Fellowship for their support; and all the Abdulkadir lab members for their valuable discussion.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.