A Double-Negative Prostate Cancer Subtype Is Vulnerable to SWI/SNF-Targeting Degrader Molecules

SWI/SNF-targeting agents interfere with a lineage-defining molecular axis in the WNT signaling–dependent, androgen receptor-negative subtype of prostate cancer, which accounts for around 10% of castration-resistant tumors.

Abstract

Proteolysis-targeting chimera (PROTAC) therapies degrading SWI/SNF ATPases interfere with androgen receptor (AR) signaling in AR-dependent castration-resistant prostate cancer (CRPC-AR). To explore the utility of SWI/SNF therapy beyond AR-sensitive CRPC, we investigated SWI-/SNF-targeting agents in AR-negative CRPC. SWI-/SNF-targeting PROTAC treatment of cell lines and organoid models reduced the viability of not only CRPC-AR but also WNT signaling–dependent AR-negative CRPC (CRPC-WNT). The CRPC-WNT subgroup represents 11% of around 400,000 cases of CRPC worldwide that die yearly. SWI/SNF ATPase SMARCA4 depletion interfered with the master transcriptional regulator TCF7L2 in CRPC-WNT. Functionally, TCF7L2 maintained proliferation via the MAPK signaling axis in this subtype of CRPC. Together, these data provide a mechanistic rationale for interventions that perturb DNA binding of the proproliferative transcription factor TCF7L2 and/or direct MAPK signaling inhibition in the CRPC-WNT subclass of advanced prostate cancer.

Significance:

SWI/SNF-targeting agents interfere with a lineage-defining molecular axis in the WNT signaling–dependent, androgen receptor-negative subtype of prostate cancer, which accounts for around 10% of castration-resistant tumors.

Graphical Abstract

Introduction

Treatment-induced shifts in cancer cell identity, known as lineage plasticity (LP), lead to the emergence of tumors that may have little-to-no resemblance to the treatment-naïve tumors. The increased and earlier use of potent targeted cancer therapies is responsible for the emergence of aggressive, “plastic,” and untreatable cancers. In prostate cancer, LP can manifest when androgen receptor (AR)–driven adenocarcinoma [AR-dependent castration-resistant prostate cancer (CRPC-AR)] differentiates into AR-negative CRPC, which lacks both canonical AR signaling and neuroendocrine (NE) differentiation markers [double-negative CRPC (DNPC)], or acquires NE features (CRPC-NE; refs. 1–3).

The epigenetic chromatin remodeling complex switch/sucrose nonfermentable (SWI/SNF) orchestrates pluripotency and differentiation in embryonic stem cells (4), indicating its potential to maintain self-renewal in cancer and modulate LP. In line with this, the SWI/SNF complex is mutated in more than 20% of cancers (5). However, in prostate cancer, genomic alterations in the SWI/SNF complex are rare. Regardless, we and others have observed dysregulation in SWI/SNF ATP-dependent helicases SMARCA2 (BRM) and SMARCA4 (BRG1) expression levels in CRPC (6, 7). In non–small cell lung cancer, alterations in SWI/SNF ATPase expression (mainly loss of SMARCA4) have led to the discovery of a synthetic lethal relationship (8, 9). In CRPC-AR, degradation of the SWI/SNF catalytic ATPase subunits (SMARCA2 and SMARCA4) compacts cis-regulatory elements bound by AR-associated transcription factors (TF), leading to a drastic decrease of prostate cancer proliferation (10).

Another critical pathway in cancer, especially during developmental processes, is the wingless and int-1 (WNT) pathway, which is partially regulated by the SWI/SNF complex (11). Aberrations in WNT signaling are especially prominent in colorectal cancer but also emerge in other cancer types, such as CRPC (12). This signaling pathway can be divided into the canonical (β-catenin dependent) and noncanonical (β-catenin independent) axis (13). In prostate cancer, genetic changes in canonical WNT pathway genes are found in up to 22% of CRPC cases, whereas noncanonical WNT signaling is also altered in advanced prostate cancer (14, 15). Recently, Tang and colleagues (16) identified a subclass of CRPC, termed CRPC-WNT, which has traits of DNPC but is enriched for mutations in WNT signaling pathway genes, accompanied by strong pathway activation through the TF TCF7L2, among others (16).

In this study, we report that targeting SMARCA2/4 downregulates this lineage-defining WNT signaling signature (16) in CRPC-WNT patient-derived organoids (PDO). Indeed, we found that SMARCA4 depletion led to the chromatin closure at TCF7L2 DNA-binding motifs and the downregulation of TCF7L2 itself in CRPC-WNT. Furthermore, we provide evidence that this downregulation of TCF7L2 is facilitated through closure of an active intragenic enhancer. By performing chromatin immunoprecipitation sequencing, we narrowed down the function of TCF7L2 in proproliferative pathways such as RAS and MEK signaling by binding to relevant gene promotors. We functionally validated that CRPC-WNT are addicted to these signaling pathways and that SMARCA2/4 degradation by A947 treatment reduces the protein levels of known MEK downstream targets.

This is in line with findings that described MEK signaling as a dependency in DNPC (1). In summary, we found TCF7L2 to be a primary driver of CRPC-WNT, which is positively regulated by SMARCA4-dependent SWI/SNF activity to drive proliferative pathways. This study strengthens the evidence that the SWI/SNF complex plays a crucial role in advanced prostate cancer and can be therapeutically exploited beyond CRPC-AR.

Materials and Methods

Cell lines and compounds

Prostate cancer cell lines (LNCaP, 22Rv1, VCaP, PC3, DU145, NCI-H660, and C4-2), other cell lines (HEK293T), and benign prostate line (RWPE-1) were purchased from ATCC and maintained according to ATCC protocols. Patient-derived CRPC organoids (WCM and MSK) were established and maintained as organoids in Matrigel drops according to the previously described protocol (17). LNCaP-AR cells were a kind gift from Dr. Sawyers and Dr. Mu (Memorial Sloan Kettering Cancer Center) and were cultured as previously described (18). All used cell lines and their phenotypes are listed in Supplementary Table S1. Cell cultures were regularly tested for Mycoplasma contamination and confirmed to be negative and kept until a maximum passage number 30. Their authenticity was confirmed by short tandem repeat profiling (tested each year, last tested in November 2025). Genentech Inc. synthesized A947, its epimer (A858), FHD-286, and AU-15330. Cobimetinib, trametinib, VL285, MBAS, and CHIR99021 were purchased from Selleckchem. BRM014, LGK974, SR18662, BGJ-398, and iCRT14 were purchased from MedChemExpress. All drugs used in this study are listed in Supplementary Table S2.

Western blotting

Whole-cell lysates were prepared in 1× cell lysis buffer (Cell Signaling Technology, 9803) supplemented with protease and phosphatase inhibitor cocktail (Thermo Fisher Scientific, 78440), and total protein was measured by the Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, 23225). An equal amount of protein was loaded in SureBlot 10% or 4% to 15% Bis-Tris Protein Gel (GenScript) and blotted. Subsequently, the nitrocellulose membrane was incubated with primary antibodies overnight in a cold room while shaking. Following incubation with horseradish peroxidase–conjugated secondary antibodies, membranes were imaged on a Vilber Fusion FX imager. Quantifications were done using ImageJ (RRID: SCR_003070). Antibodies are listed in Supplementary Table S3.

Xenograft experiment and pathologic assessment

Mice

Male NOD.Cg-Prkdcscid Il2rgtm1Wjl/Sz (NSG; RRID: BCBC_1262) mice at the age of 3 to 5 weeks were purchased from Charles River Laboratories. Mice were allowed to acclimate for 2 weeks before being used for experiments.

All animal studies were approved by the Cantonal Veterinary Ethical Committee, Switzerland (license BE35/2024). Animals were housed in ventilated cages with unrestricted access to presterilized food and fresh water. A maximum of five animals were maintained per cage on Aspen bedding. The ambient temperature was 20°C ± 2°C, and mice were kept at a constant humidity of 50% ± 10% and on a 12-hour automatic light–dark cycle.

A947 administration

Animal was restrained in injection cone for procedure and tail was warmed in water autoclaved at 37°C following disinfection of the tail before injection. Vein was visualized by slight rotation of the tail. Injection was given with a 30-gauge needle into one of the lateral veins of the mouse tail. Injection (5 µL/g) was processed slowly without aspiration. After withdrawal of the needle, injection site was carefully compressed with sterile tissue to stop eventual bleeding. Animal was checked and returned to its home cage afterward. The A947 (40 mg/kg) compound was prepared sterile in 10% hydroxypropyl-β-cyclodextrin and 50 mmol/L sodium acetate in water (pH 4) freshly on the day of injection.

In vivo experimental design

Tumor fragments (from the organoid line WCM1078) were transplanted into NSG mice. Tumor reached measurable size (60–80 mm³) after 16 days and mice were treated one time with A947 compound or vehicle intravenously into the lateral tail vein on day 19. Mice were monitored three times per week and tumor size was evaluated by digital calipering.

The volume of the tumors was calculated using the formula 4/3pi × [(sqrt(L × W))/2]3, in which L is the minor tumor axis and W is the major tumor axis. The maximal subcutaneous tumor size/burden allowed (1,000  mm³) was not exceeded in this study. Tumors and organs were harvested freshly 21 days after treatment. Fresh tissue was snap-frozen and an additional tissue sample was fixed in formalin (10%) for paraffin embedding. Paraffin-embedded tissue was cut and stained with hematoxylin and eosin stain for blinded histopathologic assessment by a board-certified veterinary pathologist (S. de Brot). Throughout the study, one animal had to be excluded because of a bacterial infection.

Immunohistochemistry

Matrigel-extracted organoids were air-dried and subsequently baked at 62°C for 25 minutes. IHC was performed on sections of formalin-fixed, paraffin-embedded organoids using a Bond automated immunostainer and the Bond Polymer Refine Detection system (Leica Microsystems, RRID: SCR_026887) by the Translational Research Unit platform, Bern (RRID: SCR_027566). The TCF7L2 antibody (Cell Signaling Technology, cat #2569, RRID: AB_2199816) was used for staining. The intensity of nuclear immunostaining was evaluated on whole slide tissue sections by a pathologist (S. de Brot) blinded to additional pathologic and clinical data.

Cell viability assay

Cells and organoids were plated in two dimensions (2D) onto Poly L–coated (LNCaP and LNCaP-AR) or Matrigel-coated (WCM and MSK lines) 96-well plates in their respective culture medium and incubated at 37°C in an atmosphere of 5% CO₂. After overnight incubation, a serial dilution of compounds was prepared and added to the plate. The cells were further incubated for 7 days, and the CellTiter-Glo 2.0 assay (Promega) was then performed according to the manufacturer’s instructions to determine cell viability. The luminescence signal from each well was acquired using the Varioskan LUX Plate Reader (Thermo Fisher Scientific, RRID: SCR_026792), and the data were analyzed using GraphPad Prism software (GraphPad software, RRID: SCR_002798).

Classification of CRPC subtypes from publicly available tumor data

Raw data in FASTQ format were obtained from the respective sources (Table 1) containing data from 45 normal prostate/primary prostate cancer samples and 29 patients with CRPC and aligned against the latest human (GRCh38) genome assembly release. Per-sample alignment and generation of feature-barcode matrices were carried out using the STARsolo algorithm (STAR version 2.7.10b, RRID: SCR_004463), tailored to the specific sequencing chemistry and the length of cell barcodes (bioRxiv 2021.05.05.442755) and unique molecular identifiers (UMI) on a case-by-case basis. We then imported the output feature-barcode matrices in an R environment (R version 4.0.2, RRID: SCR_001905) and created individual Seurat objects for each sample (Seurat package version 4.0.3, RRID: SCR_016341; refs. 19–21). A first round of quality filtering was performed through the scuttle package (version 1.8.1; ref. 22) by discarding outlier and low-quality cells after inspection of commonly used cell-level metrics (i.e., library size, UMI counts per cell, features detected per cell, and mitochondrial and ribosomal count ratio). Further doublet estimation and removal through the DoubletFinder prediction tool (version 2.0.3, RRID: SCR_018771; ref. 23) allowed us to drop unnecessary confounding technical artifacts. Thus, we merged our polished samples into a unique Seurat object. Seurat global-scaling normalization and log transformation method were applied to the complete expression matrix, followed by the selection of highly variable features and linear transformation to prepare the data for dimensional reduction. The latter scaling step also allowed us to regress out unwanted sources of heterogeneity, such as mitochondrial contamination and uneven library sizes. Thus, we determined the optimal number of principal components (PC) as the lower dimension exhibiting a cumulative percentage of variation greater than 90% to be 82. Uniform Manifold Approximation and Projection (UMAP) based on the previously selected PCs was used to reduce dimensionality and visualize the organization and clustering of cells. Specifically, we applied a graph-based unsupervised approach coupled with the Louvain clustering algorithm implemented in Seurat to generate cell clusters. We evaluated several levels of granularity through clustree (package version 0.5.0, RRID: SCR_016293; ref. 24) resolution stability analysis to accurately depict the intrinsic cellular heterogeneity. A clustering resolution of one was chosen, and marker identification was conducted by taking advantage of a hurdle model designed explicitly for single-cell RNA sequencing (scRNA-seq) data and implemented in the MAST statistical framework (RRID: SCR_016340; ref. 25), followed by a Bonferroni P value adjustment to correct for multiple testing. Only genes expressed by at least 70% of cells in the cluster, displaying a significant adjusted P value (P val_adj < 0.05) and a solid logarithmic fold change (log₂ FC > 1), were deemed as appropriate markers. After marker-based annotation of major cellular populations, we separated our collection, keeping only epithelial and malignant clusters, encompassing healthy/normal specimens of primary and castration-resistant tumors. The subsetted cells were again subjected to normalization and rescaling, and the relative UMAP was generated using 82 PCs, as before. Graph-based clustering was performed, and a granularity resolution of 0.3 was chosen, resulting in 34 clusters. In-depth annotation was assigned through marker identification. Therefore, we focused on determining which clusters could be associated with the activation of specific biological pathways by assessing the enrichment score (through the AddModuleScore function) of some recently identified signatures (16) in the CRPC setting, such as AR, NE, stem cell like (SCL), and WNT signaling. The Seurat object was then converted into an anndata object, and the Scanpy (version 1.9.5, RRID: SCR_018139; ref. 26) toolkit was used for visualization.

Table 1.

Publicly available datasets that were analyzed in this study.

Dataset ID	Number of patient samples	Sample type	Sequencing platform and chemistry	Reference (DOI)
GSE137829	6	CRPC	10x Genomics 3′ v2	10.1038/s42003-020-01476-1 (40)
GSE143791	9	CRPC	10x Genomics 3′ v2	10.1016/j.ccell.2021.09.005 (41)
GSE157703	2	Primary tumor	10x Genomics 3′ v3	10.1186/s12943-020-01264-9 (44)
GSE181294	35	Primary tumor and normal adjacent tissue	10x Genomics 3′ v2	10.1038/s41467-023-36325-2 (43)
GSE193337	8	Primary tumor and normal adjacent tissue	10x Genomics 3′ v3	10.1186/s12943-022-01597-7 (42)
GSE210358	14	CRPC	10x Genomics 3′ v3	10.1126/science.abn0478 (2)

Transfection and siRNA-mediated knockdown

ON-TARGET plus siRNA SMARTpool siRNAs against SMARCA4 (L-010431-00-0005), SMARCA2 (L-017253-00-0005), CTNNB1 (L-003482-00-0005), TCF7L2 (L-003816-00-0005), and control (D-001810-10-05) were purchased from Dharmacon. Reverse transfection was performed in six-well plates using the Lipofectamine 3000 reagent (Thermo Fisher Scientific) to the proportions of 2 μL of 20 µmol/L siRNA per well in a final volume of 2 mL. After overnight incubation, 5,000 cells were seeded as triplicates in a clear 96-well plate, and confluence was monitored using the IncuCyte S3 (RRID: SCR_023147) for up to 7 days. The remaining cells were harvested for protein extraction 96  hours after transfection.

siRNA rescue experiment

Rescue sequences for the siTools against TCF7L2 (pool of 30 siRNAs) were designed and purchased from siTOOLs Biotech. The rescue plasmid for TCF7L2 was synthesized and purchased from Atum Bio. Reverse transfection of the rescue plasmid was performed in six-well plates using the Lipofectamine 3000 reagent (Thermo Fisher Scientific) to the proportions of 5 μL of 2 μg plasmid per well in a final volume of 2 mL. After overnight incubation, cells were transfected with siTools TCF7L2 siRNA pool using the Lipofectamine RNAiMAX reagent (Thermo Fisher Scientific) to the proportions of 2 μL of 20 nmol/L siRNA per well. After overnight incubation, 5,000 cells were seeded as triplicates in a clear 96-well plate, and confluence was monitored using the IncuCyte S3 (RRID: SCR_023147) for up to 7 days. The remaining cells were harvested for protein extraction 96  hours after transfection. Sequences of siRNA and rescue sequence are shown in Supplementary Table S4.

Incucyte growth assays

2D monolayer formation

A total of 5,000 cells per well were seeded in triplicate in 96-well plates. After overnight incubation, compounds were added at indicated concentrations. Plates were monitored in the IncuCyte S3 (RRID: SCR_023147). Every 6 hours, phase object confluence (percentage area) for cell growth was measured. Growth curves were visualized using GraphPad Prism (RRID: SCR_002798).

Three-dimensional organoid formation

Using a clear 48-well plate, 20,000 cells per well were seeded in Matrigel drops. After 48-hour incubation, compounds were added at indicated concentrations. For virus transduction, lentiviral particles containing CRISPRi single-guide RNA all-in-one constructs (Supplementary Table S5) were added to organoid cells in suspension and spinfected for 4 hours before seeding in Matrigel drops. Plates were monitored in the IncuCyte SX5 (RRID: SCR_026298). Every 6 hours, the organoid object count (µm²/image) for organoid formation was measured. Growth curves were visualized using GraphPad Prism (RRID: SCR_002798).

TOPFlash reporter assay

WCM1078 cells were transduced with FOPFlash reporter (LTV-0011-4N, LipExoGen) or TOPFlash reporter (LTV-0011-4S, LipExoGen). After selection with blasticidin, cells were transduced with internal control Renilla luciferase (Rluc) lentivirus (BPS Biosciences, 79565-G). Upon selection with G418, cells were transduced with lentivirus to overexpress the empty vector (EV; GeneCopoeia, NEG-LV105), CTNNB1 (GeneCopoeia, CLP-I4822-LV105-200), or TCF7l2 (GeneCopoeia, CLP-I6388-LV105- 200-GS). After selection with puromycin, the cells were seeded as triplicates (5,000 cells/well) in a 96-well plate. Twenty-four hours later, cells were treated with either DMSO, 1 µmol/L A947, or 1 µmol/L AU-15330. Forty-eight hours later, the TOPFlash Firefly signal and the Renilla internal control signal were detected using the Dual-Glo Luciferase Assay system (Promega, E2920). Luminescence was read using the Varioskan LUX plate reader (Thermo Fisher Scientific, RRID: SCR_026792) and relative luminescence was calculated by dividing Firefly with the Rluc signal. Graphs were visualized using GraphPad Prism (RRID: SCR_002798).

scRNA-seq by SORT-seq library generation and analysis

SORT-seq was performed using Single Cell Discoveries service. Organoids were treated for 72 hours with a control epimer (A858) or active compound (A947) at 1 µmol/L, and 1 × 10e⁶ cells were harvested in PBS. Harvested cells were stained with 100 ng/mL DAPI to stain dead cells. For each treatment condition, using a cell sorter (conducted by the Flow Cytometry Core, DBMR) and the recommended settings (Single Cell Discoveries B.V.), DAPI-negative cells were sorted as single cells in 376 wells of four 384-well plates containing immersion oil, resulting in a theoretical cell number of 1,504 cells per condition. All postharvesting steps were performed at 4°C. Plates were snap-frozen on dry ice for 15 minutes and sent out for sequencing at Single Cell Discoveries B.V.

Data were analyzed using the Seurat package v.4.3.0 (RRID: SCR_016341; ref. 27). Cell quality control (QC) filtering was done using the following thresholds: nCount > 4,000, nFeature > 1,000, percent.mito < 25, and log10GenesPerUMI > 0.85. Differential gene expression analysis between clusters was done with Seurat::FindAllMarkers. Module scores were generated with Seurat::AddModuleScore. Gene set enrichment analysis (GSEA) was done with the package fgsea v.1.24.0 (RRID: SCR_020938; bioRxiv 060012) and the human gene sets from the Molecular Signatures Database (https://www.gsea-msigdb.org). Gene regulatory network analysis was done with pySCENIC v.0.12.1 (RRID: SCR_025802; ref. 28). Overall analysis was done in R v.4.2.2 (RRID: SCR_001905).

RNA sequencing library generation and processing

For bulk RNA sequencing (RNA-seq), organoids were treated with A858 or A947 (1 µmol/L) for 24 and 48 hours (three biological replicates per condition). RNA was extracted using the RNeasy Kit (Qiagen); library generation and subsequent sequencing were performed by the Clinical Genomics Lab at the University of Bern.

Sequencing reads were aligned against the human genome hg38 with STAR v.2.7.3a (RRID: SCR_004463; ref. 29). Gene counts were generated with RSEM v.1.3.2 (RRID: SCR_000262; ref. 30), for which index was generated using the GENCODE v33 primary assembly annotation. Differential gene expression analysis was done with DESeq2 v.1.34.0 (RRID: SCR_015687; ref. 31). GSEA was done with the package fgsea v.1.20.0 (RRID: SCR_020938; bioRxiv 060012) and the human gene sets from the Molecular Signatures Database (https://www.gsea-msigdb.org, RRID: SCR_016863). Analysis was done in R v.4.1.2 (RRID: SCR_001905).

TCF7L2 chromatin immunoprecipitation sequencing library generation and processing

Chromatin was prepared from two biological replicates of WCM1078 treated with A858 or A947 (1 µmol/L) for 4 hours, and chromatin immunoprecipitation sequencing (ChIP-seq) assays were then performed using an antibody against TCF7L2 (Cell Signaling Technology, cat #2569, RRID: AB_2199816). ChIP-seq data were processed using an ENCODE-DC/chip-seq-pipeline2–based workflow (https://github.com/ENCODE-DCC/chip-seq-pipeline2). Briefly, fastq files were aligned to the hg38 human genome reference using Bowtie2 (v2.2.6, RRID: SCR_016368), followed by alignment sorting (samtools v1.7, RRID: SCR_002105) of resulting bam files with filtering out of unmapped reads and keeping reads with mapping quality higher than 30. Duplicates were removed with Picard’s MarkDuplicates (v1.126) function, followed by indexation of resulting bam files with samtools (RRID: SCR_002105). For each bam file, genome coverage was computed with bedtools (v2.26.0, RRID: SCR_006646), followed by the generation of bigwig (wigToBigWig v377, RRID: SCR_007708) files. Peaks were called with macs2 (v2.2.4, RRID: SCR_013291) for each treatment sample using a pooled input alignment (.bam file) as control. Downstream analyses were performed with DiffBind v3.11.1 (RRID: SCR_012918) with default parameters, except for summits = 250 in dba.count(). dba.contrast() and dba.analyzed() were used to compute significant differential peaks with DESeq2 (RRID: SCR_015687). All ChIP-seq peaks were linked to genes with the nearest transcription start site (TSS) using chipenrich (https://www.bioconductor.org/packages/release/bioc/html/chipenrich.html.

Assay for transposase-accessible chromatin using sequencing library generation and processing

Assay for transposase-accessible chromatin using sequencing (ATAC-seq) was performed from 50,000 cryopreserved cells per condition (1 µmol/L A858 and 1 µmol/L A947, n = 3 biological replicates) treated for 4 hours and analyzed as described in a previous study (32). Briefly, 50,000 cryopreserved cells per condition were lysed for 5 minutes on ice and tagmented for 30 minutes at 37°C, followed by DNA isolation. DNA was barcoded and amplified before sequencing. All ATAC-seq peaks were linked to genes with the nearest transcriptional start site (TSS) using chipenrich (https://www.bioconductor.org/packages/release/bioc/html/chipenrich.html).

PRO-cap library generation and processing

For PRO-cap, approximately 30 million cells were processed per sample as previously described (33, 34). Library preparations for two biological replicates were performed separately. Cells were permeabilized, and run-on reactions were performed. After RNA isolation, two adapter ligations and reverse transcription were performed with custom adapters. Between adapter ligations, cap state selection reactions were carried out using a series of enzymatic steps. RNA washes, phenol:chloroform extractions, and ethanol precipitations were conducted between reactions. All steps were performed under RNase-free conditions. Libraries were sequenced on Illumina’s NovaSeq (RRID: SCR_016387) lane following PCR amplification and library clean-up. Raw sequencing data were processed as previously described (bioRxiv 2022.04.08.487666). Briefly, sequencing data were trimmed with fastp version 0.22.0 (RRID: SCR_016962) and then aligned to the human genome (hg38) concatenated with Epstein–Barr virus and human rDNA sequences (GenBank U13369.1, RRID: SCR_002760) using STAR V2.7.10b (RRID: SCR_004463). Raw alignments were filtered with samtools version 1.18 (RRID: SCR_002105) and deduplicated using umi_tools version 1.1.2 (RRID: SCR_017048). Alignments were converted to bigwig files using bedtools version 2.30.0 (RRID: SCR_006646) and kentUtils bedGraphToBigWig V2.8. Peaks were called using PINTS version 1.1.6 (35). Divergent peaks not overlapping with TSS ± 500 bp (GENCODE V37) were regarded as candidate enhancer RNAs (eRNA). GIGGLE is a genomics search engine that identifies and ranks the significance of shared genomic loci between query features and thousands of genome interval files (in our case a database of ChIP-seq experiments). A higher GIGGLE score means a stronger overlap between query features and features from the database (in our case a ChIP-seq experiment from the Cistrome database). Downstream analysis was done in R v.4.2.2 (RRID: SCR_001905). Heatmaps were generated with deepTools v.3.5.0 (RRID: SCR_016366).

Peaks were annotated with HOMER v.4.11 (http://homer.ucsd.edu/, RRID: SCR_010881). Distal peaks were defined as those peaks in known introns and intergenic regions and over 2 kb upstream or downstream from known TSS. GIGGLE scores were generated at http://dbtoolkit.cistrome.org. Analysis was done in R v.4.2.2 (RRID: SCR_001905). Graphs were generated with deepTools v.3.5.0 (RRID: SCR_016366).

TCF7L2 enhancer analysis

Hi-C and RNA-seq data for 80 metastatic CRPC biopsies had previously been generated by the Feng lab. Using the gene markers established by Tang and colleagues (16), we classified these samples into four subtypes: SCL, NE, AR dependent, and WNT signaling dependent (WNT), based on the mean log expression of the designated marker genes. Among the 80 samples, only three—DTB-135-PRO, DTB-218-BL, and DTB-130-BL—fell into the WNT category. Of these, only DTB-135-PRO had a Hi-C cis interaction depth exceeding 1 × 10⁸, making it the only viable sample for studying enhancer–promoter interactions in WNT signaling. Using the high-quality DTB-135-PRO dataset, we then applied Iterative Correction and Eigenvector decomposition (ICE) normalization, as previously described by Zhao and colleagues (36), to our Hi-C matrices at a 10 kb resolution. We examined a ±500 kb region surrounding the TCF7L2 locus and assessed Hi-C contact frequencies within A858 PRO-cap regions scoring above 10. Notably, the TCF7L2 promoter exhibited the highest contact frequency (10.1) with the suspected enhancer between chr10:113090000 to 113100000 region in the DTB-135-PRO sample.

Results

SMARCA2/4 is a vulnerability in DNPC PDOs

In prior work, we discovered that overexpression of SMARCA4 correlates with prostate cancer progression, particularly NE prostate cancer (NEPC; ref. 6). The testing of SMARCA2-/SMARCA4-targeting agents in prostate cancer had been restricted to standard cell lines, covering only a limited representation of the commonly encountered genomic landscape and progression states seen in patients with CRPC. We posited that AR-negative CRPC may also manifest sensitivity to SWI/SNF ATPase inhibition. To address this, we used a novel proteolysis-targeting chimera (PROTAC) degrader A947 that co-binds the SMARCA2/SMARCA4/PBRM1 bromodomains and the von Hippel–Lindau (VHL) ubiquitin ligase (Fig. 1A; ref. 9). This molecule has slight selectivity for SMARCA2 degradation; however, at the concentrations used in this study, it was equally potent for degradation of both ATPases within 1 hour in HEK293 cells (Fig. 1B).

Figure 1.

Drug screen identifies SMARCA2/4 as a vulnerability in DNPC cell models. A, Structure of the SMARCA2/4 PROTAC degrader A947. B, Immunoblot of indicated proteins in HEK293T cells treated with A947 (0.1 or 1 µmol/L) or DMSO over indicated time course. GAPDH represents loading control and is probed in a representative immunoblot (n = 2 independent immunoblots). C, Area under the curve (AUC) of A947 dose–response curves (see Supplementary Fig. S1A and S1B) in a panel of human-derived prostate cancer or normal cell lines after 7 days of treatment. Viability was assessed using CellTiter-Glo 2.0. Heatmap indicates gene set scores per cell model using Tang and colleagues (16) scores. Data are representative of at least n = 3 independent experiments. D, Immunoblot of indicated proteins in a panel of prostate cancer organoids representing different phenotypes. β-Actin served as a loading control and was probed in a representative immunoblot (n = 1). E, Brightfield microscopy of indicated prostate cancer organoids after 10 days of treatment with A947 (1 µmol/L) or epimer control A858 (1 µmol/L). Scale bars, 20 μm. F, Proliferation of indicated prostate cancer organoids after 10 days of treatment with A947 (at 0.25, 0.5, and 1 µmol/L) or epimer control A858 (1 µmol/L) measured by CellTiter-Glo 3D (n = 3 independent biological experiments). This represents the quantification of pictures shown in E. Data are presented as mean values ± SEM and analyzed using an unpaired Student t test. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. G, Spheroid formation of indicated prostate cancer organoids transduced with CRISPRi/Cas9 guide RNA (sgRNA) against SMARCA4 measured by live-cell imaging using IncuCyte SX5. Brightfield microscopy of indicated prostate cancer organoids at IncuCyte assay endpoint. Scale bars, 800 μm. Data are presented as mean values ± SEM and analyzed using two-way ANOVA. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. Data are representative of n = 2 independent experiments. H, UMAP plot showing disease classification (up) and relative expression of indicated signature gene scores (down) in 74 samples from six distinct prostate cancer scRNA-seq studies (2, 40–44). Side annotations indicate the AR score, NE score, SCL score, and WNT score, as determined by Tang and colleagues (16)., compared with pathology classification and molecular subtypes of each sample. FDR < 0.05. I, Percentage of indicated signatures found among all 29 CRPC-classified patient samples in H, displayed as a pie chart. na, not announced. J, Tumor growth of WCM1078 PDX subline treated with vehicle (n = 7) or 40 mg/kg A947 (n = 8) in NSG mice. Data are presented as mean values ± SEM and analyzed using two-way ANOVA (P < 0.0001).

A947 was tested on a panel of prostate cancer models, including established and organoid-derived cell lines and the nonneoplastic prostate line RWPE-1 (Fig. 1C; Supplementary Fig. S1A and S1B). Models of the four CRPC subclasses described by Tang and colleagues (16), CRPC-AR (n = 6), CRPC-WNT (n = 4), NEPC (n = 4), and SCL (CRPC-SCL; n = 7), were treated for 7 days with A947 in a dose response (Supplementary Fig. S1A). We confirmed that AR-dependent cell models are particularly sensitive to SMARCA2/4 degradation (10). In addition, we discovered that CRPC-WNT models were fully or partially responding to A947 (AUC < 250; Fig.1C; Supplementary Fig. S1A). All prostate cancer model systems tested were nonresponsive to negative control A858 epimer (SMARCA-binding control; Supplementary Fig. S1A). A947 was able to degrade all predicted targets, SMARCA2, SMARCA4, and PBRM1, in all four CRPC-WNT models (Supplementary Fig. S1C). Similar antiproliferative responses with regard to A947 were observed with SMARCA2/4 PROTAC AU-15330 and inhibitors FHD-286 and BRM014 in all four CRPC-WNT and CRPC-AR but not in other subtypes (Supplementary Fig. S2A–S2C). Next, we checked for SMARCA2, SMARCA4, and lineage-defining marker expression in selected organoids, as well as cell lines, and found the expected marker gene expression pattern, whereas the expression for SMARCA4 was overall higher than the expression of SMARCA2 in most models, also the ones that showed a response to A947 treatment (Fig. 1D).

Organoid formation was drastically reduced in CRPC-WNT organoids after 7 days of A947 treatment (Fig. 1E and F; Supplementary Fig. S2D). Assessment of growth kinetics using live-cell imaging in CRPC-WNT 2D lines MSK-PCa16 and WCM1078 showed significant reduction in cell confluence over time to a single dose of A947 (1 µmol/L) compared with control epimer A858 (Supplementary Fig. S2E).

Strikingly, competition of A947 with a free VHL ligand (VL285) rescued growth-inhibitory effect dose dependently (Supplementary Fig. S2F). Notably, levels of cleaved PARP1 (Asp214) were increased in the CRPC-WNT lines, WCM1078 and MSK-PCa16, upon treatment with 1 µmol/L A947, indicating activation of the intrinsic pathway of cell death (Supplementary Fig. S2G).

Based on these results, we wanted to test whether drugs that traditionally have been described to have an antiproliferative effect in DNPC [FGFR inhibitors (FGFRi) and KLF5 inhibitors (KLF5i)] have an effect on CRPC-WNT and/or CRPC-SCL (1, 37, 38). Indeed, we found that the CRPC-WNT model responded strongly to FGFRi BGJ-398, especially in combination with A947. This was not the case for the CRPC-SCL cell line DU145 (Supplementary Fig. S2H). When testing the KLF5i SR16882, we found that CRPC-SCL models responded stronger than CRPC-WNT (Supplementary Fig. S2I). These findings highlight, despite both CRPC-WNT and CRPC-SCL being classified as DNPC, that they are driven by different molecular mechanisms that could predict sensitivity to SMARCA2/4 inhibition.

To estimate the necessity of SMARCA2/4 activity in the CRPC-WNT, we performed individual siRNA-mediated knockdown of these two main A947 targets in the CRPC-WNT models, a CRPC-NE model, and a CRPC-SCL model. All CRPC-WNT models showed a strong growth-inhibitory effect upon SMARCA4 knockdown but only minimally responded to the knockdown of SMARCA2. The CRPC-NE model WCM154 and the CRPC-SCL model WCM155, however, were unaffected by either siRNA knockdown (Supplementary Fig. S3A and S3B). To rule out siRNA-mediated off-target effects, we elucidated the isolated effects of SMARCA4 depletion on cell growth using CRISPR/Cas9 sgSMARCA4-transduced organoids. We confirmed similar growth inhibition and cell-killing effect to A947 treatment upon SMARCA4 knockdown using CRISPRi in all four CRPC-WNT lines (Fig. 1G; Supplementary Fig. S3C and S3D). SMARCA4 inhibition has been reported to be synthetic lethal with PTEN loss in prostate cancer (39). However, only two of the four CRPC-WNT models harbor PTEN deletions, indicating that the response to SMARCA4 degradation may be independent of its PTEN status (Supplementary Fig. S4A; ref. 16). In conclusion, we identified an AR-negative subtype of CRPC that is dependent on the SWI/SNF ATPase SMARCA4 in vitro.

CRPC-WNT is a clinically relevant subset of advanced CRPC

To determine how frequently the CRPC-WNT phenotype identified by Tang and colleagues (16) is seen clinically, we used publicly available scRNA-seq data from prostate cancer cohorts [which include normal prostate/primary prostate cancer (n = 45) and samples from patients with CRPC/NEPC (n = 29) coming from six independent scRNA-seq studies; refs. 2, 40–44]. We found that the four signatures described by Tang and colleagues (16) could be identified in distinct clusters, especially the CRPC-WNT signature, which appeared in two distinct subclusters of CRPC separate from all other clusters (Fig. 1H). All remaining signatures, CRPC-AR, CRPC-NE, and CRPC-SCL, were more broadly distributed. Surprisingly, the CRPC-SCL signature was highly expressed not only in CRPC cases but also in normal and primary prostate cancer (Fig. 1H). Overall, the signatures for CRPC-AR, CRPC-WNT, CRPC-NE, and CRPC-SCL account for 38.3% (11 patients with CRPC), 11.7% (three patients with CRPC), 22.9% (seven patients with CRPC), and 22.4% (six patients with CRPC) of all CRPC cases, respectively. This underpins the clinical relevance of these signatures. These signatures could not characterize 4.7% of CRPC cases (two patients with CRPC), indicating that additional rare phenotypes of CRPC may exist (Fig. 1I). This highlights that CRPC-WNT is a clinically relevant subtype of prostate cancer for which there is no viable treatment option available. All these findings encouraged us to investigate the effect of SMARCA2/4 degradation on CRPC-WNT in vivo.

Treatment with SMARCA2/4 PROTAC leads to CRPC-WNT tumor growth delay in vivo

To assess the effect of SMARCA2/4 degradation in vivo, we generated a mouse-adapted patient-derived xenograft (PDX) subline from the WCM1078 CRPC-WNT model. As WCM1078 did not reliably grow in vivo, we had to create a subline from a single WCM1078 tumor that grew only in one of five injected mice within 6 months. Tumor cells were expanded in vitro and injected again into mice until visible tumors formed. From these tumors, we generated cryobits. The WCM1078 PDX subline cryobits showed no signs of alternative differentiation when checking the expression of relevant CRPC subtype markers, indicating that they maintained their CRPC-WNT identity (Supplementary Fig. S4A). Therefore, cryobits were subcutaneously transplanted into 20 mice. Of 20, 16 animals developed tumors; 15 were taken into the study, and the others were excluded.

A single dose of A947 treatment (40 mg/kg) or vehicle was given to eight animals or seven animals, respectively, when tumors reached a volume of 60 to 80 mm³ (Supplementary Fig. S4B). Before treatment, we checked the potential of A947 treatment to reduce mouse SMARCA4 paralog in vitro. We found that the compound is active in murine lung adenoma LA4 cells (Supplementary Fig. S4C). We observed a significant growth delay in the A947 treatment condition (n = 8) compared with vehicle control (n = 7), which aligns with the findings by Xiao and colleagues (10) in CRPC-AR (Fig. 1J; Supplementary Fig. S4D and S4E). Moreover, the mice showed no signs of aberrant behavior or reduction in body mass throughout treatment (Supplementary Fig. S4F). Histopathologic assessment indicated that all examined tissues were unobtrusive (Supplementary Fig. S4G), except for histopathologic findings in the kidney. Chronic renal interstitial fibrosis and tubular atrophy were identified in two A947-treated mice in which this change was mild (C14) to moderate (C11; Supplementary Fig. S4H). As this change is also known to occur spontaneously in laboratory mice (chronic nephropathy), it remains unknown if this lesion is related to the A947 treatment. Next, we assessed the levels of SMARCA4 in the tumors harvested at the endpoint. Quantifying the Western blot signal showed a significant average reduction in SMARCA4 protein in the A947-treated tumors compared with control despite the single-time treatment (Supplementary Fig. S4I). This indicates that A947 is highly active in vivo and can reduce CRPC-WNT tumor growth. In summary, we found that CRPC-WNT is highly responsive to a single treatment of A947 throughout a 21-day tumor growth period. These findings led us to investigate the underlying molecular mechanisms regulated by SMARCA4 in CRPC-WNT.

A lineage-defining WNT program is mitigated by SMARCA2/4 degradation in CRPC-WNT

To untangle the transcriptomic and heterogeneous changes upon treatment with A947 in CRPC-WNT, we utilized scRNA-seq by SORT-seq (45). WCM1078 and MSK-PCa16 were treated for 72 hours with 1 µmol/L A947 or control epimer A858. A total of 1,133 WCM1078 cells and 1,183 MSK-PCa16 cells in the A858-treated condition and a total of 1,184 WCM1078 cells and 997 MSK-PCa16 cells in the A947-treated condition passed QC. MSK-PCa16 demonstrated homogenous profiles with separate clusters forming based on treatment (Fig. 2A). As expected, the CRPC-WNT signature score (16) was homogenously expressed in the A858 conditions, whereas this signal was significantly reduced upon A947 treatment in MSK-PCa16 (Fig. 2B–D). These findings were comparable in WCM1078 (Fig. 2E–H). High expression in A858-treated MSK-PCa16 cells was observed for an additional WNT signature score based on colorectal cancer (CRC_WNT_score; compiled of the genes “LGR5,” “AXIN2,” “ASCL2,” “OLFM4,” “SLC12A2,” “GKN33P,” “NKD1,” and “WIF1”) although the signal was more heterogeneously expressed (Supplementary Fig. S5A). This CRC-WNT score signal was significantly reduced by A947 treatment in MSK-PCa16, in line with the reduced CRPC-WNT signature (Supplementary Fig. S5B). Comparable findings were made in A947-treated WCM1078 (Supplementary Fig. S5C and S5D).

Figure 2.

SMARCA2/4 degradation leads to strong downregulation and chromatin compaction of CRPC-WNT lineage-characterizing genes. A, UMAP plot of MSK-PCa16 organoids treated with either 1 µmol/L A858 (blue) or 1 µmol/L A947 (yellow) for 72 hours. B, UMAP plot of MSK-PCa16 organoids treated with either 1 µmol/L A858 or 1 µmol/L A947 for 72 hours displaying CRPC-WNT signature score (16). C, Violin plot of MSK-PCa16 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 72 hours displaying CRPC-WNT signature (16) score. Data analyzed using the Wilcoxon test. ****, P < 0.0001. D, Bubble plot indicative of expression levels of top 10 deregulated CRPC-WNT signature genes (16) in MSK-PCa16 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 72 hours. E, UMAP plot of WCM1078 organoids treated with 1 µmol/L A858 (blue) or 1 µmol/L A947 (yellow) for 72 hours. F, UMAP plot of WCM1078 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 72 hours displaying CRPC-WNT signature score (16). G, Violin plot of WCM1078 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 72 hours displaying CRPC-WNT signature (16) signature score. Data analyzed using the Wilcoxon test. ****, P < 0.0001. H, Bubble plot indicative of expression levels of top 10 deregulated CRPC-WNT signature genes (16) in WCM1078 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 72 hours. I, ATAC-seq read density tornado plots from WCM1078 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 4 hours (n = 3 biological replicates). J, Genome-wide changes in chromatin accessibility upon A947 treatment for 4  hours in WCM1078 organoids, along with genomic annotation of sites that gain (gained) or lose accessibility (lost) or remain unaltered (unchanged). UTR, untranslated region. K, Motifs enriched in depleted peaks from WCM1078 treated for 4 hours with 1 µmol/L A947 identified using HOMER.

However, the CRPC-SCL and CRPC-NE scores were also downregulated by A947 treatment compared with A858 treatment, but the positive clusters had low base levels compared with the CRPC-WNT–positive cluster (Supplementary Fig. S5E and S5F). This indicates that the genes that define those signatures might not be tied to canonical WNT signaling in CRPC-WNT. Therefore, we posit that the CRPC-WNT score is a refined lineage-specific signature comprised of known WNT TFs, such as TCF7, TCF7L2, and LEF1, which might have noncanonical functions.

SCENIC pathway activity analysis identified decreased activity of multiple TFs upon A947 treatment (46). In both WCM1078 and MSK-PCa16, activity of several TFs had been reduced, namely the activity of FOX, TCF/WNT, ETV, and AP-1 family members showed decreased activity (Supplementary Fig. S5G and S5H). Gene ontology analysis of the oncogene C6 signature revealed a downregulation of MAPK signaling signature terms (ERBB2_UP.V1_UP and MEK_UP.V1_UP) upon A947 treatment in WCM1078 cells (Supplementary Fig. S5I). Moreover, we found that in MSK-PCa16 scRNA-seq, the most downregulated C6 oncogene pathways were associated with WNT (LEF1_UP.V1_UP) and MAPK signatures (KRAS.600_UP.V1.UP and KRAS.KIDNEY_UP.V1_UP; Supplementary Fig. S5J). Notably, besides the two specific WNT signatures, we did not observe significant changes in other WNT pathway signatures besides one LEF-related signature (Supplementary Fig. S5I and S5J). The downregulation was confirmed for the CRPC-SCL score by bulk RNA-seq data on WCM1078 after 24 and 48 hours with A947, whereas the downregulation of the CRPC-NE score was not significant (Supplementary Fig. S6A). Related gene signatures to the ones found by scRNA-seq were found to be downregulated at earlier time points (24 and 48 hours and both overlapped; Supplementary Fig. S6B–S6D). Furthermore, bulk RNA-seq data of WCM1078 after A947 treatment showed decreased expression of a subset of the top 25 highest-ranked TFs in CRPC-WNT (16), among which, the most downregulated TFs were KLF2, TCF7L1, TCF7L2, SOX13, SOX4, RUNX3, KLF5, and LEF1 (Supplementary Fig. S6E).

Most striking, the top 10 downregulated gene signatures in bulk RNA-seq data of WCM1078 after 48 hours of treatment revealed that multiple oncogene C6 signatures, associated with MAPK–KRAS–MEK signaling, were downregulated by A947 treatment aligning with the 72-hour data (Supplementary Fig. S6F). These data validate our previous finding in the scRNA-seq analysis and point to a potential MAPK-associated proliferative axis in CRPC-WNT, which is disrupted by A947 treatment. This might be a downstream effect of the loss of the lineage-defining CRPC-WNT signature. To gain further insights into the epigenetic orchestration of this process, we examined chromatin accessibility in CRPC-WNT upon treatment with A947.

SMARCA2/4 degradation leads to closure of TCF/LEF chromatin-binding sites in CRPC-WNT

The SWI/SNF complex mediates nucleosomal DNA packaging and is actively involved in regulating gene expression of multiple programs that can be crucial for cell survival. To mechanistically exploit changes in chromatin accessibility, we profiled the changes mediated by A947 treatment using ATAC-seq. Within 4 hours, we found a near-complete loss at 3,979 sites in WCM1078 (CRPC-WNT), whereas only 80 sites were gained compared with A858 treatment (Fig. 2I). The compaction of lost sites in WCM1078 organoids was comparable with what has been found in CRPC-AR cell lines upon the treatment with SMARCA2/4 PROTAC (Supplementary Table S6; ref. 10). Overall, we saw that more than 50% of A947 treatment–compacted sites were associated with intronic and distal intergenic regions like what has been observed in CRPC-AR (Fig. 2J; ref. 10). TF motif analysis of A947-lost sites revealed lost motif accessibility of several CRPC-WNT–driving TFs, which were among others Jun-AP1, TCF, LEF, SIX, NFE, FOX, and SOX motifs (Fig. 2K; Supplementary Fig. S7A). Among the top depleted motifs, we found TCF7L2-related motifs, a WNT factor that has been identified as the most active TF in CRPC-WNT (16). Also, the motif Jun-AP1 is associated with WNT signaling and TCF7L2 as c-Jun is known to form a proproliferative complex with β-catenin and TCF7L2 in colorectal cancer (47, 48). Moreover, SIX TFs are known to directly interact with TCF7L2 and drive prostate cancer cell plasticity via WNT signaling (49–51). The fact that we see FOX motifs downregulated aligns with the findings made in CRPC-AR, indicating also the on-target selectivity of A947 (10). When performing GSEA of the ATAC-seq and ATAC-seq/RNA-seq overlap results, we found that gene sets associated with MEK signaling were found to be downregulated by A947 treatment (Supplementary Fig. S7B–S7E). These findings indicate that TCF/LEF signaling motifs are direct targets of the SWI/SNF complex, potentially regulating several downstream effectors associated with carcinogenesis and proliferation.

TCF7L2 is a dependency in CRPC-WNT

To test whether WNT signaling and TCF7L2 expression are critical for CRPC-WNT survival, we performed siRNA-mediated knockdown experiments. MSK-PCa16, which has a TCF7L2 amplification, indeed showed a growth-inhibitory effect upon siTCF7L2 transfection (Fig. 3A). This phenotype was rescued by the transfection of a transient TCF7L2 rescue construct, indicating that the effect of siRNA on cell confluence is not due to off-target effects (Fig. 3A; Supplementary Fig. S8A). In line with this, spheroid formation of other CRPC-WNT models was reduced when depleting TCF7L2 and β-catenin (CTNNB1) using siRNA (Fig. 3B; Supplementary Fig. S8B). Notably, CRPC-NE and CRPC-SCL models, which express TCF7L2, were unaffected by siTCF7L2 (Supplementary Fig. S8C). To assess potential bias from the organoid growth media, which contains R-spondin 1 (RSPO1), an activator of WNT-signaling, we measured the growth of WCM1078 and MSK-PCa16 in the absence of RSPO1 upon knockdown of β-catenin and TCF7L2. The growth phenotype with and without RSPO1 showed no significant differences when transfected with siRNA against WNT TFs, and the absence of RSPO1 only marginally affected general growth rates of these PDOs (Supplementary Fig. S8D). Indeed, a recent study has identified that RSPO1 is not the inducing factor responsible for TCF7L2 expression in WNT-driven lineage progression (52). This indicates that the WNT factors TCF7L2 and β-catenin are potentially activated via noncanonical mechanisms or are constitutively active in CRPC-WNT indicated by harbored mutations in several WNT signaling factors (16).

Figure 3.

TCF7L2 is a dependency in CRPC-WNT. A, Growth data measured by live-cell imaging (IncuCyte S3) upon transfection of indicated siRNA or plasmid (EV or rescue). Immunoblot of indicated proteins at 72 hours after transfection. GAPDH served as loading control. Data are presented as mean values ± SEM and analyzed using two-way ANOVA. ***, P < 0.001. Data are representative of n = 2 independent experiments. B, Spheroid formation measured by live-cell imaging (IncuCyte SX5) upon transfection of indicated siRNA. Data are presented as mean values ± SEM and analyzed using two-way ANOVA. *, P < 0.05; **, P < 0.01; ***, P < 0.001. Data are representative of n = 2 independent experiments. C, IHC and staining intensity of TCF7L2 on indicated organoids upon treatment with 1 µmol/L A858 or 1 µmol/L A947 for 24 hours. Violin plot from TCF7L2 staining intensity analyzed using two-way ANOVA. ****, P < 0.0001. Scale bars, 100 µm (MSK-PCa16) and 50 µm (WCM1078). OD, optical density. D, TOPFlash TCF/LEF reporter assay measured after 48 hours upon treatment with indicated drugs in WCM1078 sublines. FOPFlash served as a negative reporter control. Data are presented as mean values ± SEM after normalization to internal Renilla control and analyzed using a paired Student t test. ***, P < 0.001; ****, P < 0.0001. Data are representative of n = 2 independent experiments.

We next asked whether SMARCA2/4 degradation has direct impact on the expression of WNT signaling TFs.

IHC staining of CRPC-WNT organoids treated with 1 µmol/L of control A858 or A947 for 24 hours revealed a significant downregulation of TCF7L2 signal (Fig. 3C). We found that also the SMARCA2/4 PROTAC AU-15330 reduced levels of TCF7L2 over time (Supplementary Fig. S8E). Interestingly, protein expression of TCF7L2, TCF7, and LEF1 was reduced over time upon A947 treatment, whereas β-catenin levels were unaffected (Supplementary Fig. S8F). This led us to assess the basis of TCF7L2 downregulation upon A947 treatment by looking at changes in chromatin organization.

TCF7L2 expression is regulated through an active intragenic enhancer

As stated previously, TCF7L2-binding sites closed upon treatment with A947 and more strikingly, we observed downregulation on the protein level by IHC and immunoblot in MSK-PCa16 and WCM1078 cells within 24 hours of treatment (Fig. 3C; Supplementary Fig. S8F). Surprisingly, A947 treatment did not compact the TCF7L2 promoter loci, indicating that closure of other regulatory sites might lead to the downregulation of TCF7L2 protein. One of the earliest events in gene transcription is the activation of distal cis-regulatory enhancer regions and its associated transcription of enhancer RNA (eRNA). As changes in chromatin structure are extremely rapid upon impairment of the SWI/SNF complex, it is not surprising to see the loss of accessibility of distal regulatory regions upon SWI/SNF complex inactivation (53, 54). Thus, we hypothesized that enhancer regions of TCF7L2 could be affected by A947 treatment.

To test our hypothesis, we explored the impact of A947 treatment on eRNA expression using the nuclear run-on followed by cap selection assay (PRO-cap), which is the most sensitive method to identify active enhancers by measurement of endogenous eRNA transcription levels genome wide at base-pair resolution (35). Active enhancer loci can be precisely delineated by detecting active TSS that are dependent on the associated core promoter sequences (Supplementary Fig. S9A; ref. 55). When treating WCM1078 organoids for 24 hours with A858 or A947 (1 µmol/L), followed by PRO-cap, we found 1,069 eRNAs also known as distal peaks (±1 kb) to be downregulated (blue) whereas only two distal peaks were upregulated (red) by A947 treatment (Supplementary Fig. S9B). Next, we used the genomic search engine GIGGLE to identify and rank A947 treatment–lost genomic loci shared between publicly available genome interval files (56). These loci significantly overlapped with genomic sites bound by TFs associated with AP-1 (JUND, FOS, FOSL2, and JUN), as well as FOXA1, ETV5, and TCF7L2 among the top 20 repressed distal peaks (Supplementary Fig. S9C).

To identify potential promoter–enhancer loops regulating TCF7L2 expression, we exploited publicly available Hi-C data of 80 CRPC biopsy samples (36). From these 80 samples, one patient with the highest CRPC-WNT gene expression score was selected for further inspection. Analysis of the TCF7L2 locus showed high contact frequency of the TCF7L2 promoter (chr10:112949674–112950536) with two intragenic regions [chr10:113086750–113088000 (mean fold change over A858 = 0.2402) and chr10:113093500–113094200 (mean fold change over A858 = 0.3423)] that displayed high PRO-cap signal in WCM1078 treated with control epimer A858 (Supplementary Fig. S9D). PRO-cap signal in these two intragenic regions after A947 treatment was drastically reduced. The decrease in PRO-cap signal was accompanied with decreased ATAC-seq signal and TCF7L2 binding to these intergenic regions. These two intragenic regions were previously described as potential enhancer regions for TCF7L2 (57, 58). Based on these findings, we hypothesize that TCF7L2 regulates its own expression in patients with CRPC-WNT by binding an upstream intergenic enhancer region. Moreover, upon SMARCA2/4 degradation with A947, these potential enhancer loci showed reduced signal in PRO-cap, TCF7L2 ChIP-seq 4-hour treatment (chr10:113086750–113088000 fold change over A858 = 0.8591; chr10:113093500–113094200 fold change over A858 = 0.5549), and ATAC-seq 4-hour treatment (chr10:113086750–113088000 fold change over A858 = 0.8281, P val = 0.1186; chr10:113093500–113094200 fold change over A858 = 0.6598, P val = 0.032) assays, indicating closure of those sites (Supplementary Fig. S9D). This finding indicates that TCF7L2 expression is regulated by the SWI/SNF complex via maintenance of an intronic regulatory enhancer region in CRPC-WNT, similar to what has been reported for AR and FOXA1 in CRPC-AR (10).

TCF7L2 is not maintaining CRPC-WNT proliferation via traditional WNT signaling cues

We next assessed the effect of SMARCA2/4 degradation on TCF7L2 and WNT activity. For this, we tested whether A947 treatment interfered with transactivation of the TCF/LEF reporter TOPFlash (59). We generated stable organoid lines from WCM1078 that express the multimerized TCF-binding site TOPFlash reporter or the negative control containing mutated TCF-binding sites (FOPFlash). To know whether TCF7L2 or β-catenin overexpression (OE) could rescue the expected downregulation of reporter signal by SMARCA2/4 PROTAC treatment, we overexpressed these two factors in the TOPFlash/FOPFlash reporter organoids.

As expected, we found that both PROTACs, A947 and AU-15330, repress the TOPFlash reporter signal after 48 hours of treatment. Although β-catenin and TCF7L2 overexpression increased the reporter signal in the DMSO control condition, the PROTAC treatment–induced signal reduction could not be rescued in WCM1078 organoid lines (Fig. 3D; Supplementary Fig. S10A). As we saw that A947 treatment represses the expression of multiple TCF and LEF TFs within 24 hours, it is not surprising that overexpression of a single factor is not enough to restore the TCF/LEF reporter signal when multiple TFs remain depleted (Supplementary Fig. S8F). This indicates that SMARCA2/4 degradation leads to closure of TCF/LEF bindings sites in CRPC-WNT. As we have not observed canonical WNT signaling to be strongly affected upon SMARCA2/4 degradation but multiple WNT TFs to be downregulated by treatment, we raise the question whether these TFs drive CRPC-WNT via alternative mechanisms.

To test whether CRPC-WNT are dependent on canonical WNT signaling, we treated CRPC-WNT organoids with three WNT inhibitors that have different modes of action [LGK974 (porcupine inhibitor; ref. 60), iCRT14 (β-catenin inhibitor; ref. 61), and MSAB (β-catenin inhibitor, leading to its degradation; ref. 62)]. In addition, we treated the CRPC-NE model WCM154 and the CRPC-AR model LNCaP, which should be “WNT independent,” with these drugs. To our surprise, we found that all cell models used, including the “WNT dependent” ones, did not respond to LGK974 or iCRT14. MSAB treatment led to decreased proliferation in all cell models tested, also the “WNT-independent” ones, at approximately the same concentration, indicating potential off-target effects of this drug (Supplementary Fig. S10B). To see whether pathway activation would have a stronger effect, we tested the WNT signaling agonist CHIR99021 (GSK3β inhibitor) in the CRPC-WNT models. Despite a slight increase in proliferation (up to a concentration of 1 µmol/L) in MSK-PCa16 and WCM1262, the CRPC-WNT models were unresponsive to this agonist like we have observed with the WNT agonist RSPO1 (Supplementary Fig. S8D and S10C). Regardless, TCF7L2 overexpression enhanced growth in the WCM1078 model but did not rescue the growth delay induced by A947 and AU-15330 treatment like EV control or β-catenin–OE conditions (Supplementary Fig. S10D). The fact that forced TCF7L2 expression cannot rescue the phenotype is likely as multiple TCF7L2 binding sites are closing upon SMARCA2/4 degradation, making TCF7L2 interaction with these DNA domains impossible.

This let us to hypothesize that the CRPC-WNT subtype is not driven by the canonical WNT pathway and that TCF7L2 is hijacked to activate other pathways. Although these findings were unexpected, these data align with the fact that we do not see any canonical WNT gene sets affected by A947 treatment, despite seeing multiple WNT TFs being downregulated at the protein level over time (Supplementary Fig. S6C and S8F). Thus, we raised the question whether TCF binding sites have been reprogrammed in CRPC-WNT to drive WNT-independent pathways. Therefore, we aimed to uncover the TCF7L2-orchestrated pathways, which are affected by A947 treatment in CRPC-WNT by performing TCF7L2 ChIP-seq.

SWI/SNF ATPase degradation abrogates proliferative signaling pathways tied to TCF7L2 in CRPC-WNT

To define the TCF7L2 cistrome in CRPC-WNT, we used ChIP-seq analysis of WCM1078 organoids. In line with the chromatin closure at TCF7L2 motif sites by ATAC-seq, we found decreased TCF7L2 binding to chromatin in WCM1078 organoids upon exposure to A947 for 4 hours (Fig. 4A). A947 treatment led to the loss of 4,393 sites compared with the A858 control. Of the 4,393 lost TCF7L2 sites, 1,903 showed overlap with closing chromatin regions detected by ATAC-seq, representing a significant proportion of downregulated ChIP-seq (43%) and downregulated ATAC-seq peaks (48%). As expected, the top depleted motifs upon A947 treatment in the TCF7L2 ChIP-seq are associated with LEF and TCF7L2 (Supplementary Fig. S11A). This indicates that SMARCA2/4 degradation indeed interferes with TCF7L2 chromatin binding.

Figure 4.

TCF7L2 regulates proproliferative signatures in CRPC-WNT. A, ChIP-seq read density tornado plots from WCM1078 organoids treated with 1 µmol/L A858 or 1 µmol/L A947 for 4 hours (n = 2 biological replicates). B, Venn diagram indicating A947 treatment–lost regions from ChIP-seq, ATAC-seq, and RNA-seq data in WCM1078. GSEA was performed from 350 overlapping genes. C, Immunoblot of indicated proteins at indicated time upon treatment with 1 µmol/L A947. GAPDH served as loading control. Data are representative of n = 2 independent experiments. D, Dose–response curves with indicated drugs after measurement of proliferation with CellTiter-Glo 2.0 after 7-day treatment (n = 2 independent experiments).

GSEA of the 1,903 genes overlapping between ATAC-seq and TCF7L2 ChIP-seq revealed enrichment of depleted peaks in regions associated with proliferative genes, including previously identified MAPK-associated pathways (RAF_UP.V1_DN, EGFR_UP.V1_UP, RAF_UP.V1_UP, and MEK_UP.V1_UP; Supplementary Fig. S11B). GSEA of the intersect of RNA-seq, ATAC-seq, and ChIP-seq (350 genes) resulted in the top downregulated pathway being MEK signaling (Fig. 4B). To understand whether CRPC-WNT is dependent on the MAPK–MEK signaling axis, we checked how A947 treatment transcriptionally affects genes that define the so-called MAPK pathway activity score (MPAS; ref. 63). MPAS contains a set of MAPK downstream targets that selectively predict sensitivity to MEK inhibitors (MEKi) in multiple cancer types. Interestingly, DNPC has previously been described to be sensitive to MEKi (1). Moreover, inhibition of MEK or FGFR1 led to downregulation of the gene transcripts making up the MPAS signature (e.g., ETV4, ETV5, DUSP4, and SPRY2) in models of DNPC (1). Indeed, when checking the expression of MPAS genes from RNA-seq data after A947 treatment, we found that almost all these transcripts were downregulated in the CRPC-WNT model WCM1078 (Supplementary Fig. S12A). This was confirmed on the protein level in WCM1078 and MSK-PCa16 CRPC-WNT lines when treating the cells with A947 or AU-15330 (Fig. 4C; Supplementary Fig. S12B). As the downregulation of MPAS proteins happened only after 24 hours of A947 treatment, we postulated that this effect is downstream of TCF7L2 downregulation (which already happens within 1 hour; Supplementary Fig. S8F). Furthermore, we tested whether the MPAS is indeed predictive for sensitivity to MEK inhibition in CRPC-WNT. For this, we used MEKi cobimetinib alone or in combination with A947 or SMARCA2/4 inhibitor FHD-286 in CRPC-WNT, CRPC-NE, and CRPC-SCL lines. We found that cobimetinib alone and in combination with SMARCA2/4-interfering agents was most active in CRPC-WNT (Fig. 4D). These results were recapitulated with another MEKi, trametinib (Supplementary Fig. S12C). We tested whether TCF7L2 actively regulates the expression of MPAS genes. To address this, we wanted to know whether TCF7L2 overexpression can rescue the expression of MPAS proteins upon treatment with A947. For this, we engineered WCM1078 organoids to lentivirally overexpress TCF7L2 or an EV control. As expected, the expression of MPAS genes could be partially or fully rescued compared with EV control upon treatment with A947 for 24 hours (Supplementary Fig. S12D). Lastly, ChIP-seq of TCF7L2 confirmed binding of promoter regions of MPAS genes (Supplementary Fig. S13A–S13D). Furthermore, we found that TCF7L2 also binds the promotor of SMARCA4 but not SMARCA2 or PBRM1. This binding is reduced upon treatment with A947 (Supplementary Fig. S13E–S13G). We also found reduced binding at the SIX2 promotor in the A947 treatment condition (Supplementary Fig. S13H). SIX2 is known to be a TCF7L2 interactor, as mentioned previously, and has implications in prostate cancer lineage identity (16, 64). This indicates that TCF7L2 is participating in regulating these critical MAPK target genes in CRPC-WNT.

Thus, we conclude that the SWI/SNF complex directly shapes the cistrome for WNT signaling TF TCF7L2 in AR-negative CRPC-WNT to drive proproliferative pathways that are predictive of MEKi sensitivity.

Discussion

The standard approach to treating advanced prostate cancer has been to modulate the AR axis either through direct or indirect means (65). Drugs such as enzalutamide or abiraterone are potent AR signaling inhibitors (ARSi) used clinically and other agents, including AR degraders, are in clinical development. Resistance to ARSi therapy manifests in manifold ways (e.g., AR gene mutation, amplification, and enhancer amplification), and a subset acquires epigenetic rewiring toward AR-negative phenotypes (7).

As mentioned, AR-negative prostate cancer had previously been classified as CRPC-NE or DNPC. CRPC-NE is characterized by small cell morphology, stemness, and the expression of neuronal and NE marker genes; however, although DPNC is also AR negative, it shows no evidence of NE differentiation based on morphology or expression of classical NE markers (1, 66). From this classification emerged two novel subtypes that branch into the DNPC category: CRPC-WNT and CRPC-SCL (16). Tang and colleagues (16) suggested that CRPC-WNT is TCF/LEF TF driven, whereas CRPC-SCL is dependent on YAP/TAZ TF. Unfortunately, targeting these specific pathways directly remains a clinical challenge (67, 68). An alternative approach is to interrupt master transcriptional lineage programs by targeting TF cofactors and associated epigenetic regulators. Among these epigenetic regulators is the chromatin remodeler SWI/SNF complex, which we have previously found to be dysregulated in prostate cancer throughout disease progression and thus represents a viable therapeutic target in early but also late-stage disease (6). The SWI/SNF complex has been linked to being a predominant orchestrator of lineage-defining transcriptional programs, especially in master TF–addicted cancers (69).

A recent study in AR-dependent prostate cancer found that PROTAC degraders that target the SWI/SNF complex disrupt the enhancer and promoter looping interaction that wires supra-physiologic expression of lineage-driving oncogenes, including AR, FOXA1, and MYC (10). However, the number of AR-negative models tested in this study was limited; therefore, we examined the effect of SWI/SNF ATPase PROTAC degraders in a prostate cancer–focused screen. We utilized both AR-dependent and a broad spectrum of AR-negative prostate cancer model systems, including CRPC-NE, CRPC-WNT, and CRPC-SCL. Here, we report that VHL-dependent degraders for SWI/SNF ATPase components decrease proliferation and spheroid formation in organoids of the CRPC-WNT phenotype for which no standard-of-care treatment exists. Clinically, CRPC-WNT tumors account for around 5% to 11% of all CRPC cases (Fig. 1I; refs. 9, 10, 16). We found that the SWI/SNF ATPase SMARCA4, but not SMARCA2, is a dependency in the CRPC-WNT phenotype in vitro and in vivo.

Mechanistically, we identified that the activity of intestinal stem cell factor TCF7L2, the most active TF in CRPC-WNT (16), to be attenuated upon degradation of SMARCA2/4. To our surprise CRPC-WNT models did not respond to classical ways of WNT inhibition. This indicates that TCF7L2 is involved in maintaining a niche of DNPC but potentially via noncanonical, “nontraditional” roles of TCF/LEF signaling. In line with this, we discovered that A947 treatment reduces TCF7L2 binding to MAPK-associated gene promotors. This indicates that TCF7L2 potentially gets hijacked from its traditional role in canonical WNT signaling to assist in driving MAPK transcriptional circuits. In line with these findings, canonical WNT signaling has not been nominated as a driver of DNPC, emphasizing that TCF7L2 has different roles in the DNPC subtype termed CRPC-WNT (1). Furthermore, a link between Ras pathway activation and TCF7L2 has been reported (70). This is underpinned by the finding that MAPK signaling is a dependency in DNPC and that clinical trials with MEKi trametinib in CRPC have entered phase II (NCT02881242; refs. 1, 71). However, these trials were not biomarker based and were conducted in patients who progressed after AR-targeted therapy. Thus, based on our data, it may be beneficial to clinically assess the utility of the CRPC-WNT score as a biomarker in CRPC to predict response to SMARCA4- or MAPK-targeting therapies. Another point to consider is that we found that TCF7L2 and/or SMARCA4 levels are not necessarily predictive of response to A947 as the most responsive CRPC-WNT model WCM1078 has lower levels of TCF7L2 and SMARCA4 as the less responsive line MSK-PCa16 for example (Fig. 1D). Furthermore, we realized that certain CRPC-WNT models, e.g., MSK-PCa1, responded better to certain SMARCA2/4 inhibitors/PROTACs (FHD-268 and AU-15330) compared with others (BRM014 and A947; Supplementary Fig. S2B). Therefore, alternate expression levels and chemical properties need to be taken into consideration when utilizing SMARCA2/4-targeting agents clinically and warrant further investigation.

In summary, we nominated the SWI/SNF chromatin remodeling complex, primarily SMARCA4, as a vulnerability in DNPC classified as CRPC-WNT. Impaired maintenance of chromatin accessibility by SMARCA4-containing SWI/SNF complexes potentially blocks the binding of TCF7L2 on the chromatin, leading to reduced proproliferative pathway activity. Paralleling other studies in CRPC and small cell lung cancer, our data suggest that SWI-/SNF-targeting agents have general efficacy in cancers that are strongly driven by nuanced master transcriptional regulators (10, 72). Furthermore, we posit that MEK inhibition could be another viable approach to target CRPC-WNT and potentially other DNPC subtypes and anticipate a mechanistic connection in future work as indicated in previous studies (1, 2). We recognize that more in-depth mechanistic studies need to be conducted in this prostate cancer phenotype to fully understand the underlying complex role of TCF7L2. Lastly, we consider exploring the role of other TFs that have been affected by SMARCA2/4 degradation in CRPC-WNT in the future, such as AP-1 and FOX TFs, as they have been implicated to de-repress WNT signaling in prostate cancer (73, 74).

Supplementary Material

Supplementary Tables S1-S6

Supplementary Tables S1-S6

Supplementary Figure S1

Drug screen with SMARCA2/4 degrader

Supplementary Figure S2

CRPC-WNT is a clinically relevant subtype that can be targeted by SMARCA2/4 PROTAC degraders

Supplementary Figure S3

CRPC-WNT organoids are SMARCA4 but not SMARCA2 dependent

Supplementary Figure S4

A947-treatment is reducing CRPC-WNT tumor growth in vivo with minimal adverse effects

Supplementary Figure S5

A947-treatment consistently modulates lineage-defining CRPCWNT signature and master TFs

Supplementary Figure S6

CRPC-WNT organoids become deregulated in multiple proliferative pathways upon SMARCA2/4 degradation

Supplementary Figure S7

A947-treatment leads to decreased chromatin accessibility in CRPC-WNT

Supplementary Figure S8

CRPC-WNT is dependent on TCF7L2

Supplementary Figure S9

The TCF7L2 promoter interacts with an intragenic enhancer that is kept accessible by the SWI/SNF complex

Supplementary Figure S10

TCF7L2 is not maintaining CRPC-WNT proliferation via traditional WNT signaling cues

Supplementary Figure S11

TCF7L2 chromatin binding is reduced upon treatment with A947 in CRPC-WNT

Supplementary Figure S12

TCF7L2 regulates pro-proliferative pathways in CRPC-WNT

Supplementary Figure S13

TCF7L2 chromatin interactions

Data Availability

Sequencing data have been deposited in NCBI’s Gene Expression Omnibus (GEO) and are accessible through GEO series accession number GSE313838. The scRNA-seq datasets analyzed in this study were obtained from GEO under accession numbers GSE137829, GSE143791, GSE157703, GSE181294, GSE193337, and GSE210358 (Table 1). All other raw data generated in this study are available upon request from the corresponding author.

Authors’ Disclosures

X. Yao reports other support from Genentech Inc. outside the submitted work. S.R. Shah reports ownership of OncoVisio Inc. equity, other support from NeuScience Inc. (ownership of equity and scientific advisory board member), and other support from Third Bridge Group Limited (consultant). H. Beltran reports other support from Merck, Pfizer, Bayer, and AstraZeneca, grants and other support from Daiichi Sankyo and Novartis, and grants from Bristol Myers Squibb, Circle Pharma, and AbbVie outside the submitted work. Y. Chen reports grants from Foghorn during the conduct of the study, as well as other support from ORIC and personal fees from Belharra outside the submitted work. R.L. Yauch reports other support from Genentech outside the submitted work. M.A. Rubin reports nonfinancial support from Genentech during the conduct of the study, as well as a patent for SMARCA4 for Prostate Cancer Diagnosis and Therapeutics pending. No disclosures were reported by the other authors.

Authors’ Contributions

P. Thienger: Conceptualization, writing–original draft. I. Paassen: Investigation. X. Yao: Formal analysis. P.D. Rubin: Investigation. M. Lehner: Investigation. N. Lillis: Formal analysis. A. Benjak: Formal analysis. S.R. Shah: Investigation. A.K-Y. Leung: Formal analysis. S. de Brot: Investigation. A. Naveed: Investigation. B. Daniel: Investigation. M. Shi: Investigation. J. Tremblay: Formal analysis. J. Triscott: Investigation. G.A. Cassanmagnago: Formal analysis. M. Bolis: Formal analysis. L. Mela: Investigation. H. Beltran: Resources. Y. Chen: Resources. S. Piscuoglio: Writing–review and editing. H. Yu: Writing–review and editing. C.K.Y. Ng: Writing–review and editing. D.A. Quigley: Formal analysis. R.L. Yauch: Resources, writing–review and editing. M.A. Rubin: Conceptualization, supervision, writing–review and editing.