Loss of PBRM1 Alters Promoter Histone Modifications and Activates ALDH1A1 to Drive Renal Cell Carcinoma

The chromatin state of a cell is governed by a variety of mechanisms, including DNA methylation, histone variant incorporation, histone post-translational modification, and ATP-dependent chromatin remodeling (1). Aberrant activation or dysfunction of any of these processes can play a role in malignant transformation. SWI/SNF complexes are one class of ATP-dependent chromatin remodelers. Cancer-sequencing studies have revealed that numerous SWI/SNF subunits of the BAF and PBAF complexes are frequently mutated in human malignancies. Surveys of cancer exome–sequencing studies indicate that components of these complexes are mutated in approximately 20% of all cancers (2).

The core BAF complex consists of one of two mutually exclusive ATPase subunits: either BRG1 (SMARCA4) or BRM1 (SMARCA2), in complex with SNF5 (SMARCB1), BAF155, and BAF170 (3). BRG1-containing complexes can be further divided on the basis of their accessory subunits: ARID1A or ARID1B define the BAF complex, whereas PBRM1, ARID2, and BRD7 are found exclusively in the PBAF complex (4, 5). A third major SWI/SNF complex, termed the GBAF or non-canonical BAF complex, has also been recently identified that incorporates GLTSCR1 or GLTSCR1 L in place of an ARID protein, and contains BRD9, but lacks the core canonical BAF subunits BAF45, SNF5, and BAF57 (6).

PBRM1, also known as BAF180, Polybromo, or PB1, contains six bromodomains and two bromo-adjacent homology domains and is one of the defining subunits of the PBAF complex. We initially identified PBRM1 as a tumor suppressor in breast cancer, with a frequency of mutation of approximately 4% in cell lines and primary tumors with loss of heterozygosity (7). A subsequent exome sequencing study in clear cell renal cell carcinoma (ccRCC) revealed PBRM1 mutations in 41% of tumors, making it the second most highly mutated gene in ccRCC after VHL (8). Additional efforts by The Cancer Genome Atlas (TCGA) and others have found similar rates of PBRM1 mutation in ccRCC (9, 10). Other studies have found recurrent PBRM1 mutations in other cancer types, including pancreatic ductal adenocarcinoma (10%), intrahepatic cholangiocarcinoma (13%), gallbladder carcinoma (25%), bladder urothelial carcinoma (9%), and gastric adenocarcinoma (7%; refs. 11–14).

The molecular basis for PBRM1 tumor-suppressor function has been attributed to its ability to induce p21 and inhibit activation of S-phase gene expression, as well as its role in maintaining chromosome stability (7, 8, 15, 16). PBRM1 is also capable of recognizing activated p53 to help induce p21 and other p53-target genes (17). Other work has shown that PBRM1 loss in ccRCC amplifies the HIF pathway to promote tumorigenesis in conjunction with VHL inactivation (18). Amplification of the transcriptional outputs of HIF1, as well as STAT3, was also seen in a mouse model with dual loss of Vhl and Pbrm1 (19). This, along with other mouse studies, confirmed Pbrm1’s role as a bona fide tumor suppressor in ccRCC (20, 21). One of these studies additionally showed that Pbrm1 loss rescued cells from Vhl-loss induced replication stress to promote carcinogenesis (21). Much recent work has focused on PBRM1’s effects on the immune system and response to immunotherapy, although PBRM1’s role here remains unsettled, with conflicting findings suggesting that PBRM1 deficiency can be either immunosuppressive or pro-inflammatory (22–24).

Here, we perform epigenomic and gene expression profiling to identify and then functionally validate novel cancer-related genes and pathways regulated by PBRM1 in ccRCC. We also use biochemical methods to further explore how PBRM1 loss affects PBAF complex formation. We find that loss of PBRM1 leads to increased H3K4me3 peaks at promotors of genes associated with retinoic acid biosynthesis and signaling, leading to transcriptional activation. We functionally validate one such target gene, the aldehyde dehydrogenase (ALDH) 1A1, whose expression is consistently increased in ccRCC cell lines and primary tumors with PBRM1 loss and enhances ccRCC tumor cell growth in multiple different assays in vitro. ARID2 positively regulates ALDH1A1 in the setting of PBRM1 deficiency, whereas other core SWI/SNF components are dispensable, including the ATPase subunit BRG1. In total, this study reveals novel mechanisms of tumor suppression by PBRM1 in ccRCC.

786-O (RRID:CVCL_1051), A-704 (RRID:CVCL_1065), and ACHN (RRID:CVCL_1067) cell lines were purchased from the ATCC (which authenticates cell lines using several methods, including DNA fingerprinting) in 2010–2011. Murine embryonic fibroblasts (MEF) were generated using standard protocols from Pbrm1 fl/fl mice produced in our laboratory (2012–2014). Cell lines were clear of Mycoplasma as determined by the Lonza kit (LT07–418) within 6 months of their use. Cell lines were further authenticated in 2015 by LabCorp using a short tandem repeat method. Experiments with cancer cell lines were performed from passages 5–12. Early passage MEFs were used (<6 passages). Culturing conditions were as specified by the ATCC, unless otherwise noted.

When treated with EGF (AF-100–15, Peprotech), the cells were starved for 16 hours in the appropriate media without FBS before treatment. Erlotinib was purchased from Fisher (#50–148–625), and DEAB (dissolved in 95% ethanol) provided in the ALDEFLUOR kit (STEMCELL, RRID:SCR_013642) was used.

Full-length PBRM1 was cloned into a pBABEpuro vector. A C-terminal V5 tag was added to this plasmid using site-directed mutagenesis, and this plasmid was then used as a template to make the PBRM1 cancer mutants. The ALDH1A1-HA vector in pcDNA was purchased from Addgene (#11610, RRID:Addgene_11610).

Cells used for xenograft experiments had been split a few days before use and were healthy looking and still growing. 2 hours before injection, the cells were re-fed. They were then trypsinized, counted, washed in media, resuspended in cold media, and then combined 1:1 with cold Matrigel (Trevigen) to a concentration of 1×10⁶ per 0.2 mL and kept on ice. 6-week-old female athymic nude mice were subcutaneously injected on their ventral flanks with 1×10⁶ 786-O control and PBRM1 shRNA cells (opposite flanks of the same mouse). Tumor xenograft growth was monitored, and after the tumors were palpable (10 days), size was measured using calipers [length and width measurements were taken, and volume calculated using the formula: (length × width²)/2]. When tumor size exceeded the limitations specified by IACUC (>1,000 mm³), the mice were euthanized. These studies were approved by the IACUC, and all mice were treated humanely according to the guidelines established by IACUC.

The ALDEFLUOR assay kit was purchased from STEMCELL Technologies (RRID:SCR_013642), and the manufacturer's instructions were followed. Please see the Supplementary Information for a more detailed protocol.

Healthy, growing cells were trypsinized, counted, and washed in tumorsphere media (appropriate cell culture medium, penicillin, streptomycin, 1x B-27 Supplement (Gibco, #17504–044), 20 ng/mL EGF, and 20 ng/mL bFGF (R&D Systems, #233-FB-025). Cells were then resuspended to single-cell suspensions in tumorsphere media, and 4 × 10³ cells in 3 mL were plated into wells of ultra-low attachment 6-well plates (Costar, #3741). When drug treatments were added, they were added to the single-cell suspensions with mixing before plating. Tumorsphere formation was quantified after 3.5 days using phase contrast images at 100× of all cells in a well and analyzing the images with ImageJ software. A grouping of cells was counted as a tumorsphere if it was >50 micron in diameter, and represented a solid mass of cells, with indistinguishable cell borders. To measure tumorsphere self-renewal, the tumorspheres were passaged at this point. All the media were collected from a well, the cells were briefly spun down, and then resuspended in tumorsphere media and counted. Cells were re-plated as before at roughly the same starting density (e.g., 4 × 10³ cells in 3 mL)—if the total cell count for a well was less than 6 × 10³ cells, all the cells went back into a single well, but if the cell count was higher, the cells were divided into 2 wells. When tumorsphere formation was quantified for these cells, tumorspheres from cells that had been divided between two wells were added together. For siRNA experiments, cells were transfected with the indicated siRNA for 24 hours before they were plated for the tumorsphere assay. For the ALDH1A1 overexpression experiment, the cells were transfected for 36 hours before plating. Extra cells that were not plated were spun down, combined with 2X sample buffer, and then analyzed by Western Blot analysis to check for successful knockdown or overexpression.

Nuclear extraction was performed on 786-O and A-704 cells before the immunoprecipitation (IP) and glycerol gradient experiments. For each cell line, five 15-cm plates (80%–90% confluent) were collected at once. Each plate was washed 2× in cold PBS, and then was collected by scraping into cold PBS containing protease inhibitors (Sigma, P8340). Cells were spun at 700 × g at 4°C for 5 minutes. The pellet was then resuspended by pipetting and vortexing in 3.5 mL of cold Buffer A (10 mmol/L HEPES-KOH pH 7.9, 10 microns KCl, 0.1 microns EDTA, and 0.1 mmol/L EGTA), supplemented with fresh dithiothreitol (DTT) at 1 mmol/L and protease inhibitors. After 5 minutes, 10% Triton X-100 in Buffer A (pre-made) was added 1:20 to a final concentration of 0.5%, and the sample was vortexed again. After another 5 minutes, the sample was centrifuged at 1,000 × g at 4°C for 5 minutes. The pellet (nuclei) was then washed once by resuspending in 1 mL of cold Buffer D (20 mmol/L HEPES-KOH pH 7.9, 1 mmol/L EDTA, 1 mmol/L EGTA), supplemented with fresh 1 mmol/L DTT and protease inhibitors. The sample was centrifuged again at 1,000 × g at 4°C for 5 minutes, and the pellet was now resuspended in 0.5 mL of cold Buffer C (20 mmol/L HEPES-KOH pH 7.9, 400 mmol/L NaCl, 1 mmol/L EDTA, and 1 mmol/L EGTA), supplemented with fresh 1 mmol/L DTT, protease inhibitors, and 1 mmol/L sodium orthovanadate. The sample was vortexed, passed numerous times through an 18.5G syringe, and then incubated on ice for 5 minutes. 10% Triton X-100 in Buffer C (pre-made) was added 1:20 to a final concentration of 0.5%, and the sample was again vortexed, passed numerous times through an 18.5G syringe, and then incubated on ice for 5 minutes. The sample was then centrifuged at 14,000 rpm at 4°C for 10 minutes. The supernatant (nuclear extract) was removed, and the extraction was repeated on the pellet, beginning with resuspension in 0.5 mL of Buffer C. The second nuclear fraction was then combined with the first, and then 0.6 volumes of cold Buffer D, supplemented with fresh 1 mmol/L DTT and protease inhibitors, was added to lower the solute concentration. The nuclear extract was then filtered through a 0.22 microns filter using a 1-mL syringe, a small aliquot was set aside to quantify the protein concentration using the Bradford protein assay, and the nuclear extract was snap frozen and stored at −80°C.

IP experiments were performed on isolated nuclear extracts. 0.5 mg of nuclear extract was diluted to 1 mL total volume using cold Buffer C/D (1.0:0.6 ratio), supplemented with fresh 1 mmol/L DTT, protease inhibitors, and 1 mmol/L sodium orthovanadate. The nuclear extract was pre-cleared with 30 μL Protein G Dynabeads (Thermo Fisher Scientific) for 1 hour at 4°C. The IP step was performed overnight at 4°C using 3 μg of targeting antibody (BRG1 (G-7; Santa Cruz Biotechnology, sc-17796, RRID:AB_626762) or ARID2 (E-3; Santa Cruz Biotechnology, sc-166117, RRID:AB_2060382) or mouse IgG (Santa Cruz Biotechnology, #sc-2025, RRID:AB_737182) prebound to 30-μL Protein G Dynabeads. The beads were washed on a magnet 4× with 1 mL cold Buffer C/D, supplemented with fresh 1 mmol/L DTT, protease inhibitors, and 1 mmol/L sodium orthovanadate. After the last wash, the beads were resuspended in 60 μL of 2X sample buffer, boiled for 5 minutes, and then subjected to Western blot analysis. The V5 IP experiments in the A-704 lines were performed similarly, except 50 μL V5-agarose affinity gel (Sigma, A7345, RRID:AB_10062721) was used for the IP step, and 50 μL of protein A/G agarose beads (Santa Cruz Biotechnology, #sc-2003, RRID:AB_10201400) and 4 μg mouse IgG were used for the pre-clear step.

Glycerol gradients were made using 10% and 30% glycerol solutions, to a total volume of 11.5 mL per gradient, using a dual piston gradient maker (Jule, J17) and 13.2 mL thin-wall polypropylene tubes (Beckman Coulter, 14 × 89 mm, #331372). Glycerol solutions were made by diluting glycerol (v/v) into Buffer C/D (1.0:0.6), supplemented with fresh 1 mmol/L DTT, protease inhibitors, and 1 mmol/L sodium orthovanadate (Buffer C—20 mmol/L HEPES-KOH pH 7.9, 400 mmol/L NaCl, 1 mmol/L EDTA, 1 mmol/L EGTA; Buffer D—20 mmol/L HEPES-KOH pH 7.9, 1 mmol/L EDTA, 1 mmol/L EGTA). The fractionation was performed on isolated nuclear extracts diluted to a final concentration of 1.7 mg/mL in cold Buffer C/D (1.0:0.6), supplemented with fresh 1 mmol/L DTT, protease inhibitors, and 1 mmol/L sodium orthovanadate. 0.5 mL of the diluted nuclear extract was carefully layered on top of the glycerol gradient. The gradients were spun for 18 hours in a TH-640 swinging bucket rotor at 40,700 rpm at 4°C in an ultracentrifuge. Approximately 0.5-mL fractions (9 drops) were collected from the bottom (heavier fractions) of the tube (24 fractions total). Aliquots of the fractions were combined with 2X sample buffer and subjected to Western blot analysis.

After infection with adenovirus, MEFs were passaged 3 times, and total RNA was collected for microarray analysis using the Qiagen RNeasy kit. The Ambion WT Expression Kit (#4411973) was used to generate amplified sense-strand cDNA, and fragmentation and labeling was performed with the GeneChip WT Terminal Labeling and Controls Kit (Affymetrix, #901525). Microarray expression analysis was performed using GeneChip Mouse Gene 2.0 ST Arrays (Affymetrix, #902118) following the manufacturer's instructions.

Before RNA collection, cells were passaged and grown at the specified serum levels for 48 hours. Total RNA was collected from healthy cells that were approximately 70% confluent using the Qiagen RNeasy kit. Poly-A selection of mRNA, RNA-seq library preparation, and 100 base-pair single-end sequencing using an Illumina HiSeq machine were performed. FastQC (RRID:SCR_014583) checks were done on the raw data, and salmon (RRID:SCR_017036) was used to align the reads to the hg38 transcriptome and check for differential expression following previously described in methods (25). Salmon output (quant.sf files) was input directly into DESEQ2 (RRID:SCR_015687) for differential gene expression analysis (26). A TPM cutoff value of 2 was used to identify expressed genes. Differentially expressed genes were visualized in GSEA (RRID:SCR_003199) and Enrichr (RRID:SCR_001575) for pathway analysis (27).

Native ChIP-seq was performed according to well-established, published protocols (28). Briefly, 20–50 ×10⁶ cells were swelled and lysed using 0.1% IGEPAL, nuclei were collected by centrifugation through a sucrose gradient, and micrococcal nuclease digestion was optimized for each cell collection to primarily generate mononucleosomes, without overdigesting. For the ChIP step, 3 μg of each histone mark antibody was used, and 50 μg of chromatin was used for methyl marks, and 100 μg was used for acetyl marks. The cells were precleared by adding 30 μL of Magna ChIP Protein A+G Magnetic Beads (Millipore, #16–663) and nutating at 4°C for 20 minutes. The beads were removed, and 1% of the sample was removed and stored at −80°C as the input DNA, and the antibody incubation was performed overnight. 50 μL of Magna beads were then added, and a further 3 hours incubation at 4°C was carried out. Bead washing, processing of input DNA, and extraction and purification of DNA were performed per well-established published protocols, as above.

ChIP-seq libraries were made following well-established protocols. Briefly, 2–10 ng of DNA was end-repaired, an A-overhang was added, Illumina Truseq adapters were ligated, and DNA was run on an agarose gel and size-selected at 300–400 bp using the Qiagen Gel Extraction Kit to exclude polynucleosomes. Libraries were PCR amplified for 10–14 cycles using KAPA Biosystems HiFi PCR Master Mix. Where not indicated, all library preparation steps involved New England BioLabs enzymes. The quality and concentration of the prepared libraries were assessed on a Bioanalyzer—the amplified libraries produced sharp peaks of approximately 280 bp. 75 bp single-end sequencing was performed on an Illumina NextSeq 500 machine.

Adapter sequences were removed from reads using Cutadapt (RRID:SCR_011841). Reads were mapped to the hg19 human genome using bowtie (RRID:SCR_005476). Duplicate reads were removed using samtools (RRID:SCR_002105). Matching input control was used to call peaks. Peak calling was performed using MACS2 (RRID:SCR_013291). A P cutoff value of 10^–10 was used for peak calling. Bigwig tracks were generated using deepTools (RRID:SCR_016366) bamCoverage with RPKM normalization. H3K4me3 ChIP-seq tracks were promoter normalized before direct comparison. Blacklisted regions (Duke_Hg19SignalRepeatArtifactRegions.bed, downloaded from the Broad Institute) were excluded from called peaks using bedtools (RRID:SCR_006646). Summary plots (metagenes) and heatmaps were generated using deepTools computeMatrix with either plotProfile or plotHeatmap.

For all ATAC-seq libraries, 150,000 cells were harvested, tagmented with 5 μL Nextera Tn5 Transposes from the Nextera kit (Illumina) for 30 minutes, amplified up to 13 cycles and purified essentially as described previously (29). Purified libraries were then size selected on 2% agarose gels (150–700 bp). Libraries were sequenced on an Illumina Hi-Seq2500 (40 bp paired-end). Reads were trimmed for Illumina adapter sequences using in-house scripts and aligned to the GRCh37/hg19 using Bowtie2 (version 2.1.0) with parameters –S –X 2000. Reads that align to mtDNA, with quality value Q<30, as well as duplicated reads, were discarded using in-house scripts and Samtools. Coverage tracks were generated using deepTools bamCoverage with parameters—normalizeUsingRPKM. Tracks were further promoter normalized before direct comparison.

Statistical analysis was performed on GraphPad Prism 8 software. Statistical tests used for specific experiments are described in the figure legends. In general, a P value of <0.05 was considered significant.

Histone modification ChIP-seq, RNA-seq, ATAC-seq, and microarray datasets generated in this study are deposited at the NCBI Gene Expression Omnibus under the accession GSE102807. Any materials are available from the corresponding author upon reasonable request. RNA-seq data from previously published primary ccRCC tumors were accessed under the accession GSE86095 (30).

Further detailed methods can be found in the Supplementary Information.

To investigate PBRM1 function, we manipulated its expression in three ccRCC cell lines. We transduced 786-O cells, which are PBRM1 wild-type (WT) and VHL null, with two independent and non-overlapping shRNAs or with a non-targeting control shRNA (8). PBRM1 shRNAs reduced PBRM1 protein levels by approximately 90% (Fig. 1A). We also used shRNA to knockdown PBRM1 protein levels in the ACHN cell line (Supplementary Fig. S1A). ACHN harbors a heterozygous nonsense mutation of PBRM1 (and WT VHL); however, it has near-normal levels of WT PBRM1 protein (8). We also transduced A-704 cells, which harbor a homozygous truncating mutation of PBRM1 with mutant VHL, with either an empty-vector containing retrovirus (EV) or a retrovirus containing WT PBRM1 with a C-terminal V5 tag (WT). PBRM1 expression was stable over time in the A-704 WT cells, and no PBRM1 expression was detectable in the EV line (Fig. 1B). We next compared the growth rates of these cell lines. Under normal culture conditions (media containing 10% FBS), the growth rates between PBRM1-deficient and PBRM1-expressing cell lines were not obviously different (Supplementary Fig. S1B and S1C). However, under reduced serum conditions, PBRM1-deficient cells grew significantly faster than PBRM1-expressing control cells, as expected (Fig. 1C and D; Supplementary Fig. S1C; ref. 18).

We then performed a series of 3D growth assays to test how PBRM1 deficiency affects a hallmark of tumorigenic potential. In both 786-O and ACHN cells, PBRM1 knockdown significantly increased colony formation in soft agar (Fig. 1E; Supplementary Fig. S1D). A-704 EV and WT cells were unable to form colonies in soft agar. We next tested whether PBRM1 deficiency in 786-O cells increased tumorsphere-forming capacity, which is thought to represent a good surrogate measure of in vivo tumorigenicity. PBRM1 knockdown significantly increased the ability of 786-O cells to form free-floating tumorspheres (Fig. 1F, left), and these cells maintained a substantially higher sphere-forming capacity after passaging (Fig. 1F, right), indicating a higher self-renewal rate. Representative tumorsphere images are shown in Supplementary Fig. S1E. There were also significantly more PBRM1 shRNA cells present at the time of passage, indicating that increased tumorsphere formation was not just due to increased association of individual cells (Supplementary Fig. S1F).

To further test the 3D growth capacity of PBRM1 knockdown cells, we performed tumor xenograft experiments. 786-O PBRM1 shRNA cells grew significantly faster than control cells (Fig. 1G) and formed significantly larger tumors (Supplementary Fig. S1G). Protein extracted from the tumors confirmed that PBRM1 knockdown persisted during the xenograft experiment (Supplementary Fig. S1H).

Because prior reports have shown that PBRM1 helps regulate p21 expression and that this is an important mechanism of tumor suppression, we also tested whether PBRM1 knockdown affects p21 induction in response to doxorubicin treatment (7). In both the 786-O and ACHN lines, PBRM1 knockdown led to reduced induction of p21 in response to doxorubicin treatment (Supplementary Fig. S1I and S1J).

On the basis of the differential sensitivity to serum conditions that we observed, we hypothesized that PBRM1 may act to restrain growth factor signaling. Prior work in Drosophila melanogaster has demonstrated genetic antagonism between the PBAF homologous pbap complex and the EGFR pathway, and a more recent study using CRISPR and shRNA screens identified PBRM1 loss as attenuating EGFR inhibition by sustaining AKT signaling in a non–small cell lung cancer model (31, 32). Given these findings and the central role of the EGF pathway in cancer cell growth, we investigated the relationship between PBRM1 and EGF signaling further.

Before doing so, we created additional stable A-704 lines that expressed cancer-associated mutant versions of PBRM1 previously identified to have deleterious effects on function (p.T232P, p.A597D, and p.H1204P), and one in-frame deletion (p.M1209_E1214delMFYKKE, termed “6AAD”; ref. 8). We were unable to detect expression of the A597D mutant, and the 6AAD mutant was expressed at very low levels compared with WT (Supplementary Fig. S2A). On the other hand, the H1204P mutant expressed at comparable levels with the WT protein, whereas the T232P mutant was detectable, although at lower levels than the WT protein. IP experiments indicated that the WT, H1204P, and T232P proteins bound to other PBAF subunits (Supplementary Fig. S2B). For these reasons, we decided to use the H1204P and T232P mutant cells in follow-up experiments.

In the cell lines we created, loss of PBRM1 heightened sensitivity to EGF stimulation (Supplementary Fig. S2C–S2F). At baseline, the 786-O PBRM1 knockdown cells had higher total EGFR levels (Supplementary Fig. S2C, quantified in S2E). These cells were also more sensitive to EGF stimulation (10 ng/mL) compared with the control cells as demonstrated by increased elevation of p-EGFR (Y1143), p-AKT (T308 and S473), and p-ERK levels. At 20 ng/mL EGF, p-EGFR and p-ERK levels, but not p-AKT, remained relatively elevated in the PBRM1 knockdown lines. In A-704 cells, similar trends were evident (Supplementary Fig. S2D, quantified in Supplementary Fig. S2F). At baseline, expression of WT PBRM1 resulted in lower total EGFR levels compared with the EV control or PBRM1 mutant lines, particularly the T232P mutant. When stimulated with EGF, the WT line responded with relatively lower levels of p-EGFR (Y1173) and p-AKT (compared with all other lines at T308; only relative to the mutant lines at S473), although p-ERK levels were similar.

Next, we hypothesized that PBRM1 status would affect the growth response to EGF. In 786-O cells at very low serum, PBRM1 knockdown led to heightened growth after EGF stimulation, whereas the control cells did not respond (Supplementary Fig. S2G). In A-704 cells at very low serum, EGF stimulation produced divergent outcomes: The EV line responded with a slight growth increase, whereas EGF stimulation had a growth inhibitory effect on the WT line (Supplementary Fig. S2H), a phenomenon that is seen in some cell lines, including sometimes for ccRCC (33, 34).

In showing heightened sensitivity to EGF stimulation and increased AKT signaling with PBRM1 loss, our findings agree with the prior study in lung cancer cells (32). Although loss of PBRM1 in this lung cancer model did not lead to an increase in the total level of EGFR or downstream ERK signaling, loss of other SWI/SNF subunits in different lung cancer models has produced these changes and thus may reflect context-dependent effects (35).

We also explored whether PBRM1 knockdown influences sensitivity to EGFR inhibition using the drug erlotinib. Having established that PBRM1 knockdown has particularly large effects on 3D growth, we tested how EGFR inhibition affected colony growth in soft-agar in 786-O cells. When treated with an intermediate dose of erlotinib (1 μmol/L), the control cells almost completely lost their ability to form colonies, whereas the PBRM1 knockdown cells still formed significantly more colonies, albeit at a reduced level (Supplementary Fig. S2I). At a higher dose of erlotinib (10 μmol/L), colony growth was severely limited in both lines.

As mutations of SWI/SNF subunits have previously been linked to changes in the histone modification landscape, we next performed native chromatin IP with next-generation sequencing (NGS; ChIP-seq) probing for H3K4me3, H3K4me1, H3K27me3, and H3K9ac in both 786-O PBRM1 shRNA as well as non-targeting control shRNA cells at normal serum conditions (10% FBS). We also performed assay for transposase accessible chromatin with NGS (ATAC-seq) and RNA sequencing (RNA-seq) in the same samples for an integrative multi-omics approach to probe the impact of PBRM1 loss on the epigenome.

We observed a notable gain of H3K4me3 peaks, a marker of active and poised promoters, in the PBRM1 knockdown lines compared with the control line. Across the two PBRM1 shRNA and the non-targeting control lines, we called 16,525 high confidence H3K4me3 peaks. We identified 1,420 (8.6%) H3K4me3 peaks as gained with knockdown of PBRM1 (average shPBRM1 log₂ fold change >1; Fig. 2A and B). Conversely, only 259 (1.6%) of H3K4me3 peaks were lost with knockdown of PBRM1 (average shPBRM1 log₂ fold change <–1; Fig. 2A). We did not observe widespread changes in the other histone marks (H3K9ac, H3K4me1, and H3K27me3).

After we identified the H3K4me3 landscape as most affected by knockdown of PBRM1, we next aimed to characterize the putative upstream transcription factors involved in regulation of the 1,420 gained H3K4me3 peaks. We performed ATAC-seq analysis on the 786-O lines and identified open chromatin regions within gained H3K4me3 peaks in PBRM1 knockdown cells and performed TRAP motif analysis (36). The most highly enriched transcription factor motif was for the AP1 complex, followed by hepatocyte nuclear factor 1 beta (HNF1B), which is essential for renal development, as well as HNF 1 alpha (HNF1A; Fig. 2C; ref. 37). Motif analysis was also notable for enrichment for NFE2L2 (also known as NRF2), a major regulator of cytoprotective responses to oxidative stress, including regeneration of NADPH (Fig. 2C; ref. 38).

We further assessed whether knockdown of PBRM1 resulted in open chromatin changes that could explain the observed gained H3K4me3 peaks. We integrated the sample-matched ATAC-seq datasets and performed metagene analysis comparing open chromatin levels in PBRM1 shRNA versus control cells at the loci of gained or lost H3K4me3 peaks (Fig. 2D). We found the level of ATAC-seq reads to be comparable at these H3K4me3 peaks regardless of PBRM1 status. This finding suggests that the differences in H3K4me3 levels may be independent of changes in open chromatin, which has been noted previously for other histone modifications in cancer (39). In addition, we performed a global analysis to look for differences in the open chromatin landscape between PBRM1 shRNA and control 786-O cells. We called a total of 56,826 high confidence ATAC-seq peaks and identified 501 peaks (0.88%) as gained and 646 peaks (1.1%) as lost following knockdown of PBRM1 (absolute log₂ fold change >1). Taken together, our data suggest that PBRM1 deficiency does not profoundly impact the open chromatin landscape of ccRCC, reproducing observations made by other groups (18).

We next sought to characterize downstream transcriptional programs activated by the altered H3K4me3 peaks. Gained and lost H3K4me3 peaks correlated with mRNA expression changes in sample-matched RNA-seq datasets (Fig. 3A). We then derived a gene signature for gained H3K4me3 peaks in PBRM1 deficiency by associating the 1,420 gained peaks with genes based on the closest transcriptional start site, filtering by genes upregulated in PBRM1 shRNA cells over non-targeting controls (P_adj < 0.05, log₂ fold change > 1).

We subjected this gene signature to pathway analysis using Enrichr (40). The most significantly enriched pathway using the KEGG database was retinol metabolism (P_adj = 3.7e−4; Fig. 3B). We validated this observation using an alternative database (Reactome), similarly finding retinoic acid metabolism biosynthesis (P_adj = 0.006) and signaling (P_adj = 0.017) as the second and fourth most significantly enriched pathways, respectively, associated with gained H3K4me3 peaks in PBRM1 shRNA cells (Fig. 3C).

Cellular production of retinoic acid is tightly regulated (41). Generating the functionally active hormone retinoic acid first requires synthesis of its relatively inactive precursor, retinol, a vitamin A compound. Retinol dehydrogenases convert retinol to retinaldehyde, which becomes irreversibly converted to retinoic acid by retinaldehyde dehydrogenases. Increased retinaldehyde dehydrogenase activity induces upstream retinaldehyde reductases, which convert retinaldehyde back to retinol to maintain homeostasis, although with weak catalytic activity and modest effects on retinoic acid production in some contexts (42). Although Pbrm1 has previously been linked to retinoic acid-dependent gene activation in mouse cardiac development, to our knowledge, an association between PBRM1 loss in ccRCC and retinoic acid metabolism and signaling has not previously been described in the literature (43).

We next looked at H3K4me3 levels and expression changes at specific genes involved in retinoic acid biosynthesis and enriched in our gene signature. We observed robust increased H3K4me3 deposition at the promoter of the retinaldehyde dehydrogenase ALDH1A1 following knockdown of PBRM1 in 786-O cells (Fig. 3D). This corresponded with increased RNA-seq expression of ALDH1A1 in these cells, at both 10% and 1% FBS (Fig. 3E).

In addition, we observed increased H3K4me3 and RNA-seq expression of the retinaldehyde reductase DHRS3, as well as the retinol dehydrogenases DHRS9 and DHRS4L2 (a splice isoform of DHRS4), other genes enriched in our signature, although there was some variation in expression changes between the PBRM1 shRNA lines or the FBS conditions (Supplementary Fig. S3A–S3F; refs. 44, 45). We also identified UGT1A8, a UDP-glucuronosyltransferase involved in retinoic acid metabolism, as upregulated as part of our signature (Supplementary Fig. S3G and S3H; ref. 46). We additionally looked at the degree of open chromatin at the loci of these genes but did not see significant differences between the PBRM1 shRNA and control cells, in accordance with our global analysis (Supplementary Fig. S3I).

Having identified a possible role for PBRM1 in regulating retinoic acid biosynthesis and signaling in ccRCC, we next sought to clarify how PBRM1 loss affected one of the key enzymes in this pathway: ALDH1A1. ALDH1A1 (also known as aldehyde dehydrogenase 1 or retinaldehyde dehydrogenase 1) is one of a class of oxidizing enzymes that convert aldehydes to carboxylic acids (47). It is part of the retinoic acid metabolic pathway and irreversibly converts retinaldehyde to retinoic acid. ALDHs also protect against oxidative stress and UV damage and help catalyze the breakdown of lipid peroxides by serving as aldehyde scavengers (48). ALDH1A1 participates in hematopoietic stem cell development, white versus brown fat programming, insulin signaling, and GABA synthesis (49–52). It has also been identified as a marker of tumor-initiating cells or cancer stem cells in a variety of cancer types (53, 54).

We used qRT-PCR to confirm that ALDH1A1 mRNA levels increase with PBRM1 deficiency in the 786-O cells and decrease with PBRM1 expression in A-704 cells (Supplementary Fig. S4A). We next asked whether ALDH1A1 protein levels change accordingly. In 786-O and ACHN cells, ALDH1A1 protein increased with PBRM1 knockdown (Fig. 4A; Supplementary Fig. S4B). In A-704 cells, expression of WT PBRM1, but not the mutated forms, resulted in lower ALDH1A1 protein levels (Fig. 4B). In addition, in the immortalized breast line MCF10A, ALDH1A1 protein increased with stable knockdown of PBRM1 (Supplementary Fig. S4C), suggesting that PBRM1 regulation of ALDH1A1 may not be restricted to the malignant setting.

To further investigate this, we isolated MEFs from conditional Pbrm1 mice we generated, where Cre-mediated recombination would cause a frame-shift mutation of the Pbrm1 allele. Western blot analysis indicated effective recombination and loss of Pbrm1 expression in MEFs infected with adenovirus expressing Cre-recombinase (Supplementary Fig. S4D). We then used microarrays to perform expression analysis on three matched sets of MEFs with or without Pbrm1 expression. Aldh1a1 was among the top 50 most-altered genes in the MEFs (Supplementary Fig. S4E). We confirmed increased Aldh1a1 mRNA expression with Pbrm1 loss using qRT-PCR (Supplementary Fig. S4F).

In addition, we used the ALDEFLUOR assay to measure ALDH1-class enzyme activity in the 786-O and A-704 lines (55). The results mirrored those seen for ALDH1A1 protein levels: PBRM1 knockdown in 786-O cells increased the ALDEFLUOR-positive cell population by approximately 50%, whereas WT PBRM1 expression in A-704 cells decreased the ALDEFLUOR-positive population by approximately 33%, and the cancer mutants had no impact (Fig. 4C and D).

To examine whether ALDH1A1 expression was altered in primary ccRCC tumors with PBRM1 mutation, we used the publicly available ccRCC primary tumor expression dataset from the TCGA available on the cBioPortal for Cancer Genomics (56, 57). We divided the tumors into those with PBRM1 mutations (Mut) and those without (WT). Consistent with our cell line data, ALDH1A1 expression was significantly increased in the PBRM1 mutant setting (Supplementary Fig. S4G). Next, to discern whether these expression changes were specific to PBRM1 mutant tumors only, we divided the TCGA tumors by SETD2 and BAP1 mutation status and observed no correlation with ALDH1A1 expression levels (Supplementary Fig. S4H and S4I). We also checked for differential ALDH1A1 expression in another publicly available ccRCC expression dataset from a 2012 study that profiled BAP1 and PBRM1 mutant tumors and tumorgrafts (58). ALDH1A1 expression was significantly altered in the PBRM1 mutant setting here as well (Supplementary Fig. S4J). In addition, analyzing a previously published cohort of 10 ccRCCs, PBRM1-mutated tumors (n = 7) expressed significantly higher ALDH1A1 as compared with WT tumors (n = 3; Supplementary Fig. S4K; ref. 30).

We next investigated whether ALDH1A1 contributed to the increased 3D proliferation of PBRM1-deficient cells. Soft-agar assays in 786-O cells to which we added the ALDH-class inhibitor DEAB (which, of note, inhibits other ALDH isoenzymes in addition to ALDH1A1) revealed a dose-dependent inhibition of colony formation in PBRM1 knockdown cells (Fig. 4E; ref. 59). At the highest dose of DEAB (15 | $\rmu $ |mol/L), which is the dose used to set the negative baseline in the ALDEFLUOR assay, an equally low number of colonies could form for both control and PBRM1 knockdown cells. To show that these effects were not specific to the soft-agar assay, we also performed tumorsphere assays with or without high-dose (15 | $\rmu $ |mol/L) DEAB (Fig. 4F). The addition of DEAB almost completely abrogated the ability of 786-O cells to form tumorspheres, indicating that some level of ALDH-class activity is required for tumorsphere formation in both 786-O PBRM1 shRNA and control cells. Importantly, the addition of DEAB up to 100 | $\rmu $ |mol/L did not result in cytotoxic or cytostatic effects in 2D growth analysis for either cell line, at least at the normal serum conditions used in the experiment, suggesting that ALDH-class activity may be specifically required for anchorage independent growth (Supplementary Fig. S4L).

To demonstrate that this effect was due to ALDH1A1 inhibition specifically, we knocked down ALDH1A1 in 786-O cells using three different siRNAs (Supplementary Fig. S4M). As seen with DEAB addition, ALDH1A1 knockdown reduced the cells’ ability to form tumorspheres (Fig. 4G). The level of reduction, however, was not as complete as that seen with high-dose DEAB, suggesting either incomplete ALDH1A1 knockdown or potential compensation by other ALDH isoenzymes. There was also some variability between the ALDH1A1 siRNA oligos in their tumorsphere-reducing capacity, suggesting possible off-target effects as well. Therefore, we also assessed whether ALDH1A1 overexpression on its own could increase tumorsphere forming capacity. Compared with an EV control (pcDNA-EV), ALDH1A1 overexpression (pcDNA-ALDH1A1-HA) in 786-O control cells nearly doubled the number of tumorspheres that grew (Fig. 4H).

Given the importance of considering how remaining SWI/SNF subunits come together when one subunit is mutated, and the fact that numerous recent reports have demonstrated a synthetic lethal relationship between homologous SWI/SNF subunits, we next explored how PBRM1 deficiency affects the levels and assembly of other PBAF subunits (60–64). Western blot analysis of PBAF subunits in the 786-O and ACHN cells revealed increased ARID2 levels with PBRM1 knockdown in both shRNA lines, but not consistently increased SNF5 or BRG1 levels (Fig. 5A; Supplementary Fig. S5A). In the A-704 cells, expression of WT PBRM1 resulted in lower ARID2 protein levels, whereas levels remained stable with expression of the cancer-associated mutants (Fig. 5B). SNF5 and BRG1 levels remained stable in all A-704 lines. Transcriptional analysis using RNA-seq in the 786-O cells and qRT-PCR analysis in the 786-O and A-704 cells showed a slight trend toward increased ARID2 transcription in the PBRM1-deficient setting, but these changes were not significant (Supplementary Fig. S5B and S5C). However, in the previously mentioned publicly available ccRCC datasets, increased ARID2 mRNA levels were significantly associated with PBRM1 mutation status (Supplementary Fig. S5D). Overall, these findings suggest that increased ARID2 levels with PBRM1 deficiency may be at least partially transcriptional, but there are likely other contributing mechanisms.

We next investigated whether ARID2 protein was still assembling into a complex with other SWI/SNF components. BRG1 and ARID2 IP experiments revealed that ARID2 remained bound to BRG1 in the PBRM1-deficient setting (Fig. 5C and D; Supplementary Fig. S5E). These results are consistent with previous studies that reported PBRM1 was dispensable for the assembly of the PBAF complex (65).

These IP experiments did not preclude the possibility that ARID2 was binding to other proteins outside the PBAF complex. To determine whether this was the case, we performed glycerol gradient fractionation of nuclear extracts from the 786-O cells. On the basis of published reports, we used a 10% to 30% glycerol gradient and spun for 18 hours to ensure separation of lower molecular weight complexes (60, 63). Analysis of fractions 2–24 (out of 24 total fractions) revealed that SWI/SNF subunits, including PBRM1 and ARID2, eluted only in the heavier fractions (Supplementary Fig. S5F). No ARDI2 was found in lighter fractions. We then looked at elution patterns at higher resolution in the heavier fractions only (Fig. 5E). In the control cells, ARID2 and PBRM1 could be found in fractions 1–5, but peaked in fractions 1–2. BRG1 and SNF5 were found mostly in fractions 4–6. These results imply that the majority of ARID2 and PBRM1 is found outside the PBAF complex in the control cells. After PBRM1 knockdown, ARID2 shifted to lighter fractions, now peaking in fractions 4–5. BRG1 and SNF also shifted to slightly lighter fractions and could now be found in fractions 4–7, peaking in fraction 5. Of note, there were small differences in the elution patterns between the two PBRM1 shRNAs lines, with more SNF5 remaining in fraction 4 and ARID2 eluting in a broader range of fractions in PBRM1 shRNA #2 cells. The elution patterns of ARID2, BRG1, and SNF5 are quantified in Supplementary Fig. S5G. Overall, these results support the IP findings and suggest that after PBRM1 knockdown, ARID2 is bound in a PBAF-like complex that also contains BRG1 and SNF5. This complex elutes in slightly lighter fractions, perhaps reflecting the loss of PBRM1 inclusion.

On the basis of its domain structure containing six bromodomains capable of binding acetylated histones and the fact that it is one of the defining subunits of the PBAF complex, PBRM1 has been proposed as a targeting SWI/SNF subunit, helping direct the complex to genomic loci. In addition, many of the PBRM1 mutations are in the bromodomains. Because of this and our finding that a PBAF-like complex containing ARID2 remained intact after PBRM1 knockdown, we attempted to probe genomic occupancy of ARID2 with and without PBRM1 knockdown. However, we were unable to detect high-quality ARID2 ChIP-seq peaks using four commercially available antibodies in 786-O cells.

On the basis of our IP and glycerol gradient fractionation experiments, we next asked whether ARID2 contributed to ALDH1A1 regulation in the setting of PBRM1 deficiency. When we knocked down ARID2 using siRNAs in 786-O PBRM1 knockdown cells, ALDH1A1 protein expression declined significantly (Fig. 6A; Supplementary Fig. S6A). PBRM1 protein levels also declined after ARID2 knockdown, confirming previous findings that ARID2 is required for PBAF-complex stability (65). The ALDH1A1 changes were confirmed to occur at the mRNA level (Supplementary Fig. S6B). In A-704 EV cells, ARID2 knockdown likewise decreased ALDH1A1 protein expression but not as completely as in the 786-O cells (Supplementary Fig. S6C). We next tested whether ARID2 knockdown affected clonal expansion via the tumorsphere assay. With ARID2 knockdown, the tumorsphere forming capacity significantly declined for 786-O cells (Fig. 6B; Supplementary Fig. S6D). The results resembled the effects of knocking down or inhibiting ALDH1A1.

We also tested whether other PBAF subunits were required for ALDH1A1 upregulation with PBRM1 deficiency. We knocked down SNF5 or BRG1 using siRNAs in 786-O cells (Fig. 6C and D). However, unlike ARID2, SNF5 and BRG1 were dispensable for the high levels of ALDH1A1 expression in the PBRM1 knockdown line. We also performed the SNF5 knockdown in the A-704 lines (Supplementary Fig. S6E). In the EV cells, SNF5 knockdown led to an increase in ALDH1A1 expression, whereas in the WT cells, no ALDH1A1 expression was detectable. Although the ATPase subunit BRM has not previously been reported to bind in a complex with ARID2, considering the seeming dispensability of BRG1, we wondered whether a subunit switch of BRM for BRG1 was occurring with PBRM1 deficiency. We tested this by knocking down BRM in 786-O cells, but again saw no changes in ALDH1A1 protein levels in the setting of PBRM1 knockdown (Supplementary Fig. S6F). To test whether either ATPase subunit was required, we also performed the double knockdown of BRM and BRG1, but still saw no changes in ALDH1A1 protein expression with PBRM1 knockdown (Supplementary Fig. S6G). As BRG1 knockdown did not affect ALDH1A1 protein levels in the PBRM1 knockdown cells, we next tested how BRG1 knockdown affected tumorsphere-forming capacity. Although BRG1 knockdown did not decrease tumorsphere formation, as expected given the ALDH1A1 stability in this setting, BRG1 siRNA #5 increased tumorsphere formation compared with the control siRNA in the control shRNA cells, whereas BRG1 siRNA #1 had a similar effect in the PBRM1 shRNA cells, suggesting possible off-targets effects from these oligos as well as ALDH1A1-independent effects on tumorsphere formation (Fig. 6E; Supplementary Fig. S6H).

On the basis of our biochemical and ChIP-seq findings, we propose that loss of PBRM1 results in changes to the epigenomic landscape, mostly in the form of gained H3K4me3 peaks. These gained H3K4me3 peaks were notably associated with increased gene expression and enriched for a retinoic acid biosynthesis and signaling gene signature in the setting of PBRM1 deficiency. We leveraged this signature to identify PBRM1’s regulation of ALDH1A1, an ALDH that irreversibly converts retinaldehyde to retinoic acid, and found that ALDH1A1 can promote tumorigenesis in ccRCC.

In contrast with H3K4me3, other histone marks, including H3K9ac, H3K4me1, and H3K27me3, remain largely unchanged. Our integrated ATAC-seq analysis indicates that the H3K4me3-gained peaks are independent of changes in open chromatin. In fact, we did not see profound global changes in the open chromatin landscape with PBRM1 deficiency in 786-O cells, in line with a prior report looking at nucleosome positioning using MNase-seq (18). Instead, the gain of H3K4me3 peaks may be due to the recruitment of activating transcription factors or other co-factors. Loss of PBRM1 could serve to activate methyltransferases such as MLL1 that methylate H3K4, or block demethylases such as KDM5A-C that demethylate H3K4.

In addition, with PBRM1 depletion, we saw few changes in the H3K27me3 landscape. This suggests that PBRM1 loss does not lead to heightened EZH2 methyltransferase activity that has been seen in other cancer types with SWI/SNF subunit mutations (66, 67). However, some studies have shown that EZH2 is essential in PBRM1-mutant cancers, but that these and other SWI/SNF-mutant cancers may be primarily dependent on a non-catalytic role of EZH2 in stabilizing the PRC2 complex (68). Our results do not preclude the possible role of non-canonical EZH2 and PRC2 functioning.

Indeed, as mentioned previously, the loss of PBRM1 did not significantly impact the open chromatin landscape, suggesting that PBRM1 may be dispensable for this role in the PBAF complex. After loss of PBRM1, ARID2 still associates in a PBAF-like complex, in agreement with previous reports on PBAF complex assembly. Compatible with these assembly dynamics, we found that although ARID2 is required for PBRM1 stability, PBRM1 is dispensable for ARID2 expression. In fact, we found that ARID2 levels increase with PBRM1 deficiency, and that slightly increased ARID2 transcription may only partially account for this change, potentially reflecting post-translational stabilization. Fractionation experiments revealed that in control 786-O cells, the majority of ARID2 and PBRM1 elute in the heaviest fractions, separate from the major peaks of BRG1 and SNF5, and thus presumably outside of the major population of the SWI/SNF complex. This is in line with a previous study of SWI/SNF fractionation patterns in mouse brain nuclear extracts that also found most of the PBRM1 eluting apart from other PBAF components in the heaviest fractions (63). The authors conjectured that the PBRM1 in these fractions could be part of the mitotic machinery, as PBRM1 had previously been found to localize to the kinetochore during mitosis (69). In our cells, with PBRM1 knockdown, we saw a shift in the ARID2 elution pattern away from these heavy fractions and into lighter fractions also containing BRG1 and SNF5, but not into other fractions devoid of BRG1 and SNF5. These findings suggest that PBRM1 deficiency is compensated by increased or stabilized ARID2 expression and possibly increased ARID2 association with SWI/SNF components.

Interestingly, however, although ARID2 is required for the upregulation of ALDH1A1 seen with PBRM1 deficiency and the corresponding increase in tumorigenic potential, the core PBAF subunits SNF5 and BRG1 seem dispensable. Although subtle changes in the level of open chromatin could be occurring and having impact, the seeming stability of chromatin accessibility and dispensability of the actual chromatin remodeling subunits (BRG1 and BRM) suggest that the ARID2 complex devoid of PBRM1 could be acting at least partially in more novel, uncharacterized ways to affect transcription. Of note, this does not preclude the necessity of chromatin remodeling or BRG1 for other oncogenic effects of PBRM1 loss in ccRCC, as has been seen in other studies (18). Alternatively, the regulation of ALDH1A1 expression by ARID2 in the PBRM1-deficient setting could be explained by loss of proper functioning of the higher molecular-weight complex containing ARID2 (and potentially PBRM1) seen in the 786-O control cells. This model would also account for the chromatin accessibility and SNF, BRG1, and BRM findings. Further work is needed to better elucidate this mechanism and characterize the composition and function of the higher molecular-weight ARID2-containing complex.

We decided to further investigate PBRM1 regulation of ALDH1A1 expression. Numerous studies have identified ALDH1A1 as a marker of tumor-initiating cells, although it is unclear whether ALDH1A1 promotes tumorigenicity and how it does so. Not much is known about the regulation of ALDH1A1 expression in cancer, although recent studies have posited upregulation by β-catenin or Smad4-mediated repression in prostate cancer cells (70, 71). Another study has described inhibition of ALDH1A1 by post-translational acetylation in breast cancer cells, which is reversed by SIRT2 action downstream of NOTCH signaling (72). Our results indicate a novel method of regulation by PBRM1. The association between PBRM1 status and ALDH1A1 levels was confirmed in multiple primary tumor datasets of ccRCC. The association was also seen in the immortalized breast line MCF10A, as well as MEFs where Pbrm1 was deleted by Cre recombinase, suggesting that PBRM1 may play a role in retinoic acid homeostasis outside of the malignant setting. In addition, our results suggest that ALDH1A1 is more than just a marker of tumor-initiating cells but can directly increase the tumorigenic potential of cells. Regulation of retinoic acid biosynthesis and signaling may also have important implications for PBRM1’s role in modifying the tumor immune microenvironment. A recent study found that increased retinoic acid production by tumors because of IL-13 stimulation from the tumor microenvironment generates a more immunosuppressive microenvironment that promotes tumor immune evasion and tumor growth (73). Regulation of ALDH1A1 by PBRM1 and ARID2 may represent an alternative or additional mechanism of modulating retinoic acid levels in the tumor microenvironment, although additional work here is needed. Overall, our data offer the intriguing possibility that targeting ALDH1A1 could be beneficial for patients with PBRM1-mutated cancers.

D.A. Schoenfeld reports grants from National Institutes of Health during the conduct of the study. R. Rabadan is a founder of Genotwin, an infectious disease diagnostic company; consults for Arquimea Research; as well as is a part of the SAB of Aimedbio. R. Parsons reports grants from NIH during the conduct of the study; and other support from Therapten and personal fees from Cullinan outside the submitted work. No disclosures were reported by the other authors.

D.A. Schoenfeld: Conceptualization, formal analysis, investigation, methodology, writing–original draft, writing–review and editing. R. Zhou: Formal analysis, investigation, writing–review and editing. S. Zairis: Formal analysis, investigation. W. Su: Investigation. N. Steinbach: Methodology. D. Mathur: Investigation, methodology. A. Bansal: Investigation. A.L. Zachem: Investigation. B. Tavarez: Investigation. D. Hasson: Software, methodology. E. Bernstein: Supervision. R. Rabadan: Supervision. R. Parsons: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.

We thank Dr. R. Sachidanandam and the Department of Oncological Sciences Core for performing the ChIP-seq sequencing and offering advice and guidance on analysis. We would also think to thank Asif Chowdhury from the Bernstein laboratory for technical advice on native histone modification ChIP-seq. We would also like to acknowledge the Transgenic Mouse Core at Columbia University for helping make the Pbrm1 fl/fl mouse, and the Genomics Core Facility at the Department of Genetics and Genomic Sciences for performing the RNA-seq library preparation and sequencing. We thank the Steven Johnson laboratory for providing the ALDH1A1-HA vector through Addgene. Finally, we would like to acknowledge Drs. J. Kitajewski, W. Gu, A. Ferrando, and T. Maniatis for their helpful advice and guidance. This work was supported by the National Institutes of Health, grants R01 CA082783 (to R. Parsons), T32 GM 7367–36 (to D. Schoenfeld), U54 CA193313 (to R. Rabadan), 5TL1 TR000082 (to S. Zairis), F30CA243207 (to R. Zhou), R35CA220491, R01CA230854, and P30CA196521 (to R. Parsons), and R01CA154683 and R01CA218024 (to E. Bernstein).

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.