DROSHA regulates mesenchymal gene expression in Wilms tumor

Abstract

Wilms tumor, the most common pediatric kidney cancer, resembles embryonic renal progenitors. Currently, there are no ways to therapeutically target Wilms tumor driver mutations, such as in the microRNA processing gene DROSHA. Here we used a “multi-omics” approach to define the effects of DROSHA mutation in Wilms tumor. We categorized Wilms tumor mutations into four mutational subclasses with unique transcriptional effects: microRNA processing, MYCN activation, chromatin remodeling, and kidney developmental factors. In particular, we find that DROSHA mutations are correlated with de-repressing microRNA target genes that regulate differentiation and proliferation and a self-renewing, mesenchymal state. We model these findings by inhibiting DROSHA expression in a Wilms tumor cell line, which led to upregulation of the cell cycle regulator cyclin D2 (CCND2). Furthermore, we observed that DROSHA mutations in Wilms tumor and DROSHA silencing in vitro were associated with a mesenchymal state with aberrations in redox metabolism. Accordingly, we demonstrate that Wilms tumor cells lacking microRNAs are sensitized to ferroptotic cell death through inhibition of glutathione peroxidase 4 (GPX4), the enzyme that detoxifies lipid peroxides.

INTRODUCTION

Wilms tumor is the most common pediatric kidney cancer and one of the most common solid tumors of childhood¹. Because of their histological similarity to fetal kidneys, Wilms tumors are thought to arise from developmental derangements in embryonic renal progenitors. Typical “triphasic” Wilms tumors contain three cell types that resemble fetal kidney: blastema, stroma, and epithelia. The blastemal mesenchyme is thought to give rise to the other two compartments². Although cure rates for Wilms tumor exceed 90%, multimodal therapy can cause significant long-term therapy-related toxicity³. On the other hand, there are limited therapeutic options for those with high-risk features, and there are currently no ways to therapeutically target driver mutations in Wilms tumor.

For example, Wilms tumors commonly exhibit loss-of-function mutations in microRNA (miRNA) processing enzymes, including DROSHA, DICER1, and DGCR8^4–8. Loss of miRNA processing leads to de-repression of miRNA target genes, such as LIN28A/B and PLAG1^9,10. These mutations are often associated with Wilms tumors that bear abundant blastema, and loss of miRNA-mediated gene expression is thought to lock cells into an undifferentiated state^11,12. Other Wilms tumors are driven by mutations in WT1, SIX1, SIX2, and CTNNB1— transcription factors that regulate the specification and differentiation of embryonic nephron progenitors^5–7. Abnormal methylation at the 11p15 H19/IGF2 locus is common in Wilms tumor but is insufficient for tumorigenesis, as it is also seen in precursor lesions known as nephrogenic rests¹³.

In other undifferentiated mesenchymal cancers, therapy resistance can be overcome by inducing a death process known as ferroptosis^14–16. Ferroptosis is a regulated cell death process that occurs when polyunsaturated fatty acids (PUFAs) in the plasma membrane undergo lipid peroxidation, which sets off a chain reaction of free-radical damage^17–20. To avoid ferroptosis, lipid peroxides must be detoxified by one key enzyme: glutathione peroxidase 4 (GPX4)²¹. Agents that induce ferroptosis by interfering with GPX4 activity such as RSL3 are under investigation as potential cancer therapeutics. Acyl-CoA synthetase long-chain family member 4 (ACSL4) performs the first step in bringing polyunsaturated fatty acids to the plasma membrane and is thought to be a rate-limiting step in execution of ferroptosis. However, mutational features that regulate ACSL4 expression or activity are unknown, and little is known about ferroptosis in Wilms tumor.

Here, we sought to describe the molecular effects of tumor-driving mutations in 94 Wilms tumors. We performed an integrated multi-omics analysis with targeted capture and whole-genome sequencing, small RNA sequencing, whole transcriptome sequencing, and metabolomic profiling. We found that Wilms tumor mutations clustered into four mutational subgroups with unique, direct transcriptional effects. In particular, miRNA loss leads to de-repression of miRNA target genes that regulate proliferation and differentiation of mesenchymal renal progenitors. These cells overexpress ACSL4, which leads to enhanced sensitivity to the ferroptosis inducer RSL3. In summary, we find that Wilms tumors lacking miRNAs are sensitized to ferroptosis.

MATERIALS AND METHODS

Wilms tumors from UT Southwestern and The Children’s Hospital at Westmead

Wilms tumor samples were obtained from biospecimen repositories at UT Southwestern and The Children’s Hospital at Westmead. Samples in these repositories were collected at the time of either tumor resection or biopsy, and residual tissue was stored after obtaining written informed consent in accordance with Declaration of Helsinki guidelines. The study was approved by the Institutional Review Board of the University of Texas Southwestern Medical Center. Genomic DNA was prepared from tumor samples using the DNeasy tissue Kit (Qiagen cat 69504, Germantown, MD). Small and large RNA were prepared from UTSW samples using the miRNeasy tissue kit (Qiagen 217004). DNA and RNA from The Children’s Hospital at Westmead were prepared using the DNeasy and RNeasy kits (Qiagen cat 69504 and 74004).

Sequencing and variant calling

We chose the panel of genes to be sequenced based on previously reported recurrent mutations in Wilms Tumors ^4,22–25. Custom capture probes were designed to cover all exons and splice junctions. A total of 94 samples passed quality check. Library preparation and sequencing (100-bp paired-end, ≥1500x depth) were performed at BGI America (Cambridge, MA). The analysis workflow was based on Genome Analysis Toolkit (GATK, v3.8–0, RRID:SCR_001876)^26,27 best practices. Sequencing quality was evaluated using NGS QC Toolkit (v2.3.3, RRID:SCR_005461)²⁸, and high-quality reads were mapped to the UCSC hg38 reference genome using Burrows-Wheeler Aligner (BWA, v0.7.15a, RRID:SCR_010910)²⁹. Picard (v2.12.0, RRID:SCR_006525) (https://broadinstitute.github.io/picard) was used to remove PCR duplicates, and GATK was used to recalibrate base qualities. Calling variants and genotyping were performed using HaplotypeCaller and the variant calls were filtered using the following criteria: QD (Variant Confidence/Quality by Depth) < 2, FS (Phred-scaled p-value using Fisher’s exact test to detect strand bias) > 60, MQ (RMS Mapping Quality) < 40, DP (approximate read depth) < 3, GQ (Genotype Quality) < 7. The variants were annotated using a custom Perl script (https://github.com/jiwoongbio/Annomen) with human transcripts, proteins, and variations (RefSeq and dbSNP build 150).

Exome capture was performed at the UT Southwestern McDermott Next-Generation Sequencing Core using SureSelect Human All Exon v4+UTRs (Agilent Technologies). Sequencing was performed with a HiSeq 2000 instrument (Illumina) with 100-bp paired-end reads. Raw reads were mapped to reference genome (hg38) using BWA ²⁹. Duplicates were removed using Picard (http://picard.sourceforge.net/). Local realignment and base quality recalibration were performed using default parameters with the GATK pipeline³⁰. Matched tumor-normal BAM files were used as inputs to identify somatic SNVs and indels using the GATK pipeline.

Genomic DNA from 16 tumor-normal pairs was analyzed by WGS at the UT Southwestern McDermott Next-Generation Sequencing Core. The WGS data were aligned to the reference genome (hg19) using BWA v.0.7 and sorted by SAMtools (v.1.977, RRID:SCR_002105) and then lifted to hg38. PCR duplicates were removed by Picard (http://picard.sourceforge.net/). By taking matched normal samples as background, tumor-specific copy number alterations and LOH were called by the Somatic Copy-number and Heterozygosity ALteration Estimation (SCHALE)³¹ algorithm using default parameters.

Small RNA sequencing, whole transcriptome sequencing and pri-microRNA processing analysis

Small RNA libraries were prepared from 32 Wilms tumor samples with adequate RNA. For small RNA, libraries were prepared and sequenced at DNALink, Inc. using the NEBNext Small RNA Library Prep Kit for Illumina, and reads were sequenced on the Illumina NextSeq 500 platform. Trimmed reads were mapped to miRbase using miRDeep2 and normalized to the spike-in control. Differential expression analysis was performed using DESeq2.

Fifty-six tumors and three normal kidney samples underwent whole-transcriptome sequencing. For whole transcriptome sequencing, libraries were prepared using the TruSeq Stranded Total RNA Library Prep Kit (Illumina) and sequenced on the Illumina NovaSeq 6000 platform at DNALink, Inc. 100-bp paired-end reads were assessed for quality, and reads were mapped using CASAVA (Illumina). The generated FASTQ files were aligned by Bowtie2³² and TopHat2³³ to the hg38 human assembly. Cufflinks^34,35 was used to assemble and estimate the relative abundances of transcripts at the gene and transcript levels.

For whole transcriptome sequencing of cultured cells, libraries were prepared using the TruSeq Stranded Total RNA Library Prep Kit (Illumina). These were sequenced on the Illumina NextSeq 2000, 75-bp single-end with read depth 35–45 million each. For small RNA sequencing of these cells, 3 µg total RNA was supplemented with a spike-in control for normalization (Qiagen 331535). Libraries were prepared and barcoded with Illumina TruSeq Small RNA Sample Preparation Kit (Illumina, San Diego CA), pooled and sequenced single-end read in NextSeq 500 System with 130 M with averaging at 15 million reads per sample.

Differential expression analysis was performed using DESeq2. Specifically, a multivariate linear model was built using each mutational subclass as an independent categorical variable, allowing calculation of the contribution of each mutational subclass to expression of each gene. From these results, the Wald statistic of each gene (calculated as log2-fold-change divided by its standard error) was used for GSEA, using the fgsea package³⁶. Unless otherwise stated, gene sets were derived from Molecular Signatures Database (MSigDB v7.1)³⁷. Fusion genes were identified by STAR-Fusion (https://github.com/STAR-Fusion/STAR-Fusion) and MapSplice³⁸, separately. Only fusion transcripts identified by both algorithms were used in this study.

To study processing of pri-microRNAs, we examined whole-transcriptome sequencing reads that aligned to pri-miRNAs of commonly expressed miRNAs (i.e., microRNAs that were detectable in at least 24 of 32 tumors) whose pri-microRNA cleavage sites (i.e., the first and last position seen in pre-microRNAs) were defined in miRbase. To determine the overall efficiency of 5p and 3p cleavage, we calculated the number of aligned reads around generic 5p and 3p cleavage sites. Ratio of 5p to 3p cleavage efficiency was defined as the ratio of the read depth at the 5p cleavage site to the read depth at the 3p cleavage site.

Statistical analysis

Data analyses were conducted in R v.3.3.2 and Python v.2.7.11 or v.3.5.4. FDR-adjusted p values were based on the Benjamini–Hochberg method. P < 0.05 was considered to be statistically significant.

Re-analysis of TARGET Wilms tumors

Previously reported mutations in Wilms tumor were downloaded from the Catalogue of Somatic Mutations In Cancer (COSMIC v88, https://cancer.sanger.ac.uk/cosmic). These data include mutations previously reported in TARGET Wilms tumors. Recurrently mutated positions in DROSHA were identified using this combined source. Next, reported mutations from TARGET tumors were integrated with processed, anonymized Wilms tumor TARGET copy number data and clinical annotations from the TARGET data matrix (https://ocg.cancer.gov/programs/target/data-matrix). At any given gene, tumors were designated as having copy number loss or gain when log2 copy number was < −0.3 or > +0.3, respectively. These copy number changes were used for mutational classification: copy-number gain of MYCN (MYCN); copy-number loss of REST (chromatin remodeling); and copy-number loss of WT1, AMER1, or RERE (kidney development).

Wilms tumor RNA-seq TARGET data was processed from raw fastq files downloaded from dbGaP. Reads were aligned to the genome (hg38) using Hisat2 (v2.1.0, RRID:SCR_015530) and assembled using StringTie (v.1.3.2d, RRID:SCR_016323)³⁹ based on the GENCODE v26 reference annotation. Differential gene expression analysis was performed using DESeq2 (v1.36.0, RRID:SCR_000154)⁴⁰. The four mutational classes (miRNA, MYCN, chromatin remodeling, and kidney development) were used as independent covariates in DESeq2 calculations.

Wilms tumor metabolomics

For metabolomic analysis 34 tumor samples were flash frozen in liquid nitrogen and stored in −80°C prior to preparation. To extract metabolites, 1 ml 80% (v/v) methanol was added to 50–100 mg of the tumor. The sample was homogenized using an ultrasonicator, then 200 µl of this homogenate was diluted with 800 µl ice-cold 80% (v/v) methanol. The sample was vortexed for 1 minute, centrifuged at 20,000×g for 15 minutes at 4°C. The pellet was lysed using RIPA lysis buffer (Sigma) supplemented with protease and phosphatase inhibitor (Thermo Scientific) as previously described⁹, and quantified for normalization. The supernatant was concentrated using as SpeedVac. Screening for 151 metabolites was done via Liquid Chromatography-Mass Spectrometry and analyzed as previously described^41–44.

Cell culture

The Wilms tumor lung metastasis cell line WiT49 (RRID: CVCL_0583) was a gift from Dr. Sharon Plon’s laboratory. It was maintained in DMEM with 1000 mg/L glucose, 4 mM L-glutamine, 1 mM pyruvate (Gibco 11995065) supplemented with antibiotic-antimycotic (Gibco 15240062) and fetal bovine serum (Sigma F2442) to 10% (v/v) final concentration. Cells were split 1:4 twice weekly and grown in a humidified incubator at 37°C and 5% CO₂. Short tandem repeats (STR) genotype of WiT49 cells were verified upon initial receipt and annually, with the latest on February 2024 at the UT Southwestern McDermott Sequencing Core. Mycoplasma contamination is ruled out upon receipt and biannually (latest negative test, February 2024; Lonza LT07–318).

Knockdown of DROSHA and DICER1 in WiT49 cells was performed using CRISPR interference (CRISPRi) was as described⁴⁵. Guide RNA sequences for the genes of interest and non-targeting controls were adapted from Horlbeck et al.⁴⁶ and cloned into the vector backbone pLV hU6-sgRNA hUbC-dCas9-KRAB-T2a-Puro (Addgene plasmid # 71236; RRID:Addgene_71236)⁴⁵. Transfected cells were selected in media supplemented with 0.5 µg/mL Puromycin (Thermo Fisher J61278). All downstream applications were performed on cells no more than 9 passages since being revived from cryopreservation.

RSL-3 treatment in vitro

To quantify inhibitory activity, 1000 cells were seeded per well in 96-well plates and incubated overnight. RSL-3 (MedChemExpress HY-100218A) dissolved in DMSO was serially diluted into 10 different concentrations and added to the cells in at least 4 replicates, along with 4 cell-free and 4 DMSO-only control wells. All wells had a final volume of 100 µl. Seventy-two hours later, cell density was assayed by adding 20 µl 15 µg/ml sterile resazurin dissolved in PBS (pH=7.4) to each well for 1–4 hours, then read for its fluorescence using a microplate reader with 550 nm excitation and 590 nm emission. Fluorescence viability readings were normalized to readings from the cell-free and DMSO-only wells. IC50s were calculated using GraphPad Prism 10 (RRID:SCR_002798) through nonlinear regression [inhibitor] vs. response variable slope (four parameters) with the following constraints: Bottom = 0, Top =1, IC50>0, HillSlope <0.

To measure lipid peroxidation, cells were pulse-treated with 800 nM RSL-3 for 2 hours and then stained with 10 µM BODIPY™ 581/591 C11 (Invitrogen D3861) for 30 minutes at 37°C. The cells were resuspended in 3% BSA in PBS prior to analysis by flow cytometry. Fluorescence at 510 nm was quantified using FlowJo™ Software (BD Life Sciences).

qPCR

Total RNA was extracted from sub-confluent cells using the miRNeasy kit with DNase I digestion (Qiagen Cat #217004 and #79254). The RT2 HT First Strand Kit (Qiagen Cat #330411) was used for cDNA synthesis, and qPCR was performed using iTaq™ Universal SYBR® Green Supermix (Bio-Rad cat 1725125) with primers listed here: 18s (GTAACCCGTTGAACCCCATT, CCATCCAATCGGTAGTAGCG); ACSL4 (CCTCCGATTGAAATCACAGAAGA, TTGCCATAGCGTTTTTAAATGTCC); DICER1 (TCGAGCCTCCATTGTTGGTC, TGGTCATCCAGTTCGCCAAT); DROSHA (CCCGAGAGCCTTTTATAGGTTG, TCTGTATCCTTCACATCCCCG); GAPDH (CGGAGTCAACGGATTTGGTC, ACCAGAGTTAAAAGCAGCCC).

For miRNA quantification, reverse transcription was performed with Multiscribe Reverse transcriptase (Invitrogen 4311235). qPCR was performed with TaqMan Universal Master Mix II (Applied Biosystems 4440047) and TaqMan™ Small RNA Assays [Applied Biosystems 4427975 Assay IDs: (U6) 1973, (mir-16) 391, (mir-20a) 580]. All qPCR reactions were done in at least 4 replicates and analyzed using Student’s t test.

Transfection

To alter miRNA levels in vitro, we used mimics (control: Dharmacon CN-001000-01-05; hsa-miR-16-5p: Dharmacon C-300483-03; hsa-miR-20a-5p: Invitrogen MC10057) and inhibitors (Control: Dharmacon 1H-001005-05; hsa-miR-16-5p: Dharmacon IH-300483-05; hsa-miR-20a-5p: Dharmacon IH-300491-05-0005). These were transfected into WiT49 at 50–70% confluency using Opti-MEM™ Reduced Serum Medium (Gibco 31985062) and Lipofectamine™ RNAiMAX Transfection reagent (Invitrogen 13778150) according to manufacturer recommendations. Cells were collected for quantification of miRNA and target genes of interest 48 hours after transfection.

Western blots

Whole-cell protein lysates were obtained from cells at 50–70% confluency using RIPA lysis buffer (Sigma R0278) supplemented with protease and phosphatase inhibitor (Thermo Scientific A32961) as previously described⁹. Western blots were performed at minimum twice to confirm reproducibility. Primary antibodies and dilutions are listed here: ACSL4 (1:3000, Invitrogen, cat. PA5–27137, RRID:AB_2544613); CCND2 (1:1000, Cell Signaling Technology, cat. 3741, RRID:AB_2070685); DICER1 (1:3000, Cell Signaling Technology, cat. 5362, RRID:AB_10692484); DROSHA (1:3000, Cell Signaling Technology, cat. 3364, RRID:AB_2238644); GAPDH (1:3000, Cell Signaling Technology, cat. 97166, RRID:AB_2756824); tubulin (1:3000, Cell Signaling Technology, cat. 3873, RRID:AB_1904178).

Data Availability

Sequencing data is available at dbGAP accession number phs003559.v1.p1.

RESULTS

Wilms tumor mutational classes correlate with developmental arrest and CCND2 overexpression

To study the effect of mutations in the microRNA processing pathway, we collected Wilms tumor specimens from UT Southwestern and The Children’s Hospital at Westmead in Australia. We performed targeted deep sequencing of 67 genes in 102 Wilms tumor samples from 94 patients (Figure 1A; Supplementary Figure S1A; Supplementary Tables S1 and S2). We also performed whole-exome sequencing (WES) on 47 cases and whole-genome sequencing (WGS) on 16 tumor-normal pairs, including two tumors that were not analyzed by WES (Supplementary Figure S1B). Lastly, we performed small RNA sequencing on 32 tumors and whole transcriptome sequencing (RNA-seq) on 56 tumors.

Figure 1. Mutations observed in Wilms tumor.

(A) Mutations detected in targeted sequencing panel of Wilms tumor, classified into mutational subclasses. (B) Copy number changes detected by WGS in Wilms tumors. (C) CCND2 expression in Wilms tumors characterized by WGS. Tumors with gain of chromosome 12 are highlighted in red. (D) Mutational subclasses of Wilms tumor predict transcriptional changes.

We identified 265 protein-altering mutations by WES (Supplementary Tables S3 and S4, Supplementary Figure S1A). All 52 mutations detected in the targeted sequencing panel were validated in WES. Consistent with previous reports, tumors exhibited a low mutation rate (0.15 non-silent mutations per Mb)^4–7 (Supplementary Figure S2A). The most common somatic SNVs were C>T transitions, a pattern that differentiates Wilms tumor from renal cell carcinoma (Supplementary Figure S2B, Supplementary Figure S2C). We also identified copy number changes by WGS (Figure 1B, Supplementary Figure S1B). Consistent with previous reports, chromosome 11p frequently underwent copy-neutral loss of heterozygosity, loss of 11p, or focal loss of 11p13, the region surrounding WT1. Similarly, chromosomes 1q, 8, and 12 were recurrently amplified, and chromosomes 8 and 12 were usually amplified together^47–49. Chromosome 8 encodes for several genes that may drive tumor formation, including MYC and PLAG1, and chromosome 12 encodes for both CCND2 and its binding partner, CDK4. One tumor exhibited amplification of chromosome 12 without amplification of chromosome 8, but it also harbored a MAX mutation that could phenocopy MYC amplification (Supplementary Figure S1B). CCND2 overexpression has been previously implicated as a Wilms tumor driver⁵⁰. We found that tumors with copy-number gain of chromosome 12 were also among those that expressed high levels of CCND2 (Figure 1C). Lastly, in one tumor we also detected an EIF5-LIN28B fusion by RNA-seq (Supplementary Table S5). Similar translocations involving 6q21, where LIN28B resides, have been described in Wilms tumor^47,51–53.

Altogether, we categorized these driver mutations into four classes: miRNA processing, MYCN/MAX, chromatin remodeling, and kidney developmental factors. To validate this mutational classification, we next examined the transcriptome of each category. For example, by comparing tumors with miRNA processing mutations to those without mutations, we verified that miRNA target genes were enriched in DROSHA-mutant Wilms tumors (Figure 1D, Supplementary Figure S3A). Similarly, MYCN-mutant tumors demonstrated enrichment for both sets of MYC target genes. In the chromatin remodeling mutational cohort, the transcriptional repressor REST (also known as NRSF) was frequently mutated, and this mutational cohort exhibited de-repression of REST target genes. Lastly, mutations in either kidney developmental factors or miRNA processing mutations correlate with an expression pattern that resembles that of “self-renewing nephron progenitor cells”, the most undifferentiated nephrogenic population in the developing kidney⁵⁴.

We also examined whether correlations between mutational class and transcriptomic signature were replicated in an independent dataset from the Therapeutically Applicable Research to Generate Effective Targets (TARGET) project⁷. We re-categorized TARGET tumors into the same mutational classes based on their published mutations and copy number changes (Supplementary Figure S4A). Once again, these mutational categories had expected effects on gene expression (Supplementary Figure S4B-C). In summary, Wilms tumors have a relatively few mutations, but these mutations can lead to widespread effects on the transcriptome.

Defective microRNA processing in mutant Wilms tumors

Having observed that microRNA pathway mutations correlate with overexpression of microRNA target genes, we next examined how these mutations affect microRNA processing. We specifically examined mutations in DROSHA, one of the two most commonly mutated genes in Wilms tumor. In addition to p.E1147K, the most common DROSHA mutation, we compiled previously reported Wilms tumor mutations throughout DROSHA from the COSMIC database (Figure 2A). DROSHA has tandem ribonuclease (RNase) III domains, termed RNase IIIa and IIIb, which cleave the 3p and 5p sides of pre-microRNA hairpins, respectively. The DROSHA missense mutations seen in Wilms tumor usually affect negatively-charged metal-binding residues (Asp or Glu) of either RNase III domain. Mutations also affect Gln 1187, which also protrudes into the RNase IIIb metal-binding pocket (Supplementary Figure S5).

Figure 2. DROSHA mutations impair microRNA processing.

(A) Known Wilms tumor mutations in DROSHA prior to our study. (B) Patterns of abnormal processing of pri-microRNAs by RIIIb- or RIIIa-mutant DROSHA. (C) Small RNA sequencing in Wilms tumors with or without mutations in microRNA processing genes, normalized to a spike-in control. (D) Reads mapping to pri-microRNA regions in whole-transcriptome sequencing. (E) Ratio of cleavage at 5p versus 3p sides of pre-microRNAs. (F) GSEA from both datasets showing enrichment of self-replicating nephron progenitor cell (NPC) signature and epithelial mesenchymal transition signature in miRNA mutational class. (G) Expression of CCND2 in our panel and in TARGET Wilms tumors (p=0.008 and 0.011, respectively).

Using in vitro modeling, we previously showed that the most common heterozygous DROSHA mutation (p.E1147K) impairs microRNA production in a dominant-negative manner⁴. However, to our knowledge, no studies have demonstrated a decreased microRNA abundance in DROSHA-mutant Wilms tumors. In our small RNA sequencing, normalized to a spike-in control, we observed that tumors with mutations in the miRNA processing pathway exhibited significant miRNA depletion (Figure 2B–2C, Supplementary Table S6). Although these mutations are heterozygous, leaving one allele intact, the levels of canonical (DROSHA-dependent) miRNAs are depleted further than half. Levels of “mirtrons”, which are cleaved by the mRNA splicing machinery, and other DROSHA-independent miRNAs⁵⁵ remain unchanged (Figure 2C). We next examined pri-microRNA cleavage using whole-transcriptome sequencing. Because pre-microRNAs (60–80 nt) and mature microRNAs (–20 nt) are too short to be detected by standard protocols, tumors with normal miRNA processing have few reads aligning to pre-miRNA regions. Thus, in tumors with intact DROSHA, we observed few reads aligning to pre-miRNAs, surrounded by an accumulation of reads on each side (Figure 2D). In tumors with RIIIb mutations, many more reads span the 5p than the 3p cleavage sites, representing improperly processed pri-miRNAs. Conversely, in an RIIIa-mutant tumor, reads spanning the 3p cleavage site far outnumbered reads spanning the 5p site. To quantify the efficiency of cleavage on each side of the pri-miRNA hairpin across all tumors, we calculated the ratio of the read depth at the 5p and 3p cleavage sites (Figure 2E). Tumors with wild-type DROSHA had ratios near 1, whereas the three RIIIb-mutant tumors had ratios greater than 2. Conversely, the only tumor with a ratio less than 0.5 was the RIIIa-mutant tumor. These ratios suggest that in the heterozygous state, mutant DROSHA competes with wild-type DROSHA for pri-microRNA substrates.

To understand how mutant DROSHA contributes to tumorigenesis, we next explored enrichment of other gene sets. We noted that microRNA production mutations correlated with enrichment of the primitive nephron progenitor signature and the “Epithelial-mesenchymal transition” hallmark gene set, signifying a more mesenchymal phenotype (Figure 1D, 2F, Supplementary Figure S3B). Loss of miRNA processing may contribute to Wilms tumor formation by arresting “blastema” in an undifferentiated, mesenchymal state¹². In fact, CCND2, a microRNA target gene which is also a marker of undifferentiated NPCs is overexpressed in both cohorts (Figure 2G). Together, these findings suggest that miRNA loss contributes to Wilms tumor formation by arresting nephrogenic development and driving proliferation partially through de-repression of CCND2.

DROSHA regulates growth and differentiation signals in vitro

Among these three genes, we focused on CCND2 because its overexpression has been previously associated with Wilms tumor⁵⁰, but it be regulated by let-7 and other microRNAs post-transcriptionally⁵⁶. Thus, to understand how microRNAs affect gene expression in an isogenic context, we next silenced DROSHA or DICER1 with CRISPR interference^57,58 (CRISPRi) using two independent sgRNAs each in the anaplastic Wilms tumor cell line WiT49, which lacks mutations in microRNA processing genes (Figure 3A). Like a dominant-negative mutation, this led to dramatic but incomplete loss of microRNAs (Figure 3B). Using RNA-seq, we found that impairing microRNA production in WiT49 had similar effects to DROSHA mutations in Wilms tumor samples. Specifically, DROSHA knockdown led to an enrichment for microRNA target genes, the self-renewing NPC signature, and a more mesenchymal phenotype (Figure 3C). At the individual gene level, CCND2 is the single most overexpressed gene when DROSHA expression is inhibited (Figure 3A, 3D). In other words, loss of microRNA processing in vitro recapitulates the pro-growth and anti-differentiation signals in Wilms tumor.

Figure 3. Impairment of microRNA processing leads to CCND2 upregulation.

(A) Western blots demonstrating knockdown of DROSHA or DICER1, using CRISPR interference, using two independent sgRNAs. Non-targeting controls (sgNTC-2 and sgNTC-3) used as negative controls. (B) Total microRNA reads from small RNA sequencing of DROSHA-knockdown cells, normalized to spike-in control. (C) GSEA of DROSHA-knockdown cells demonstrates enrichment for microRNA target genes, self-renewing NPC signature, and epithelial-mesenchymal transition signature. (D) Volcano plot of protein-coding genes expressed in total RNA-seq comparing sgDROSHA cells to sgNTC WiT49 cells, with magnitude change (log2 fold change) on the horizontal axis and statistical significance (-log10(p)) on the vertical axis.

DROSHA regulates sensitivity to ferroptosis

Lastly, we noted that miRNA processing mutations were also associated with dysregulation of several metabolic pathways by GSEA (Supplementary Figure 3B). Thus, we next performed metabolomic profiling on 34 Wilms tumor samples and compared metabolomic profiles of tumors with mutations in microRNA processing genes to tumors without such mutations (Supplementary Figure S6, Supplementary Table S7). One of the most downregulated metabolites in mutant tumors were tetrahydrobiopterin (BH₄), an antioxidant that protects cells from reactive oxygen damage and ferroptotic cell death^59–61. Furthermore, DROSHA mutation enhances mesenchymal phenotype, and mesenchymal cancer cells, particularly in the kidney, are known to be at particular risk for ferroptosis^15,62–65. This is due in part to higher levels of acyl-CoA synthetase long-chain family member 4 (ACSL4), the enzyme that regulates incorporation of PUFAs into the plasma membrane. PUFAs are the key substrate for lipid peroxidation, and ACSL4 expression alone directly correlates with sensitivity to ferroptosis inducers in multiple cancers^62,63.

We noted that DROSHA silencing increased ACSL4 expression by RNA-seq and confirmed higher levels by qPCR and Western blot in both DROSHA and DICER1 knockdowns (Figure 4A-4B). We next investigated whether ACSL4 upregulation could be attributed to lack of an individual microRNA. Among microRNAs that could regulate ACSL4 in Wilms tumor, we focused on miR-16–5p and miR-20a-5p because they are both highly expressed across Wilms tumors and have well-conserved binding sites in ACSL4. We found that inhibitors of miR-20a-5p consistently increased ACSL4 in control cells, while mimics of miR-20a-5p consistently repressed ACSL4 in DROSHA-knockdown cells (Figure 4C). These results suggest that miR-20a-5p is a key regulator of ACSL4 expression in Wilms tumor.

Figure 4. DROSHA/DICER1 regulate ferroptosis.

(A) ACSL qPCR in DROSHA-knockdown WiT49 cells. *p < 0.05. (B) Western blots demonstrating upregulation of ACSL4 in DROSHA/DICER1-knockdown WiT49 cells. DICER1 and GAPDH blots on the right are the same as those shown in Figure 3A. (C) Expression of ACSL4 in WiT49 sgNTC cells transfected with microRNA inhibitors or WiT49 sgDROSHA cells transfected with microRNA mimics. * p < 0.05. (D) Inducibility of lipid peroxidation in modified WiT49, measured by flow cytometry positivity for BODIPY C11 (p = 0.11 comparing log-odds ratios of positivity for sgDROSHA/sgDICER1 versus sgNTC). (E) Sensitivity to RSL-3 cytotoxicity in modified WiT49.

Next, we used a flow cytometric assay to measure lipid peroxidation levels in these cells. We observed no differences in lipid peroxidation at baseline, but upon exposure to the GPX4 inhibitor RSL-3, DROSHA-knockdown and DICER1-knockdown cells exhibited higher levels of lipid peroxidation (Figure 4D). Furthermore, RSL-3 was significantly more toxic to DROSHA- or DICER1-knockdown cells than control cells (Figure 4E).

DISCUSSION

In this study, we performed DNA, RNA, small RNA, and metabolomic profiling on a large panel of Wilms tumors. These allowed us to show that Wilms tumor mutations affect at least four classes of genes: miRNA processing, MYCN activation, chromatin remodeling, and kidney development. These mutations appear to arrest embryonic renal progenitors in an undifferentiated, self-renewing state, and DROSHA-mutant tumors are particularly mesenchymal. We then used an in vitro system to describe how miRNA loss affects gene expression by regulating CCND2 expression. Lastly, we show that the mesenchymal state induced by DROSHA mutation also sensitizes these cells to GPX4 inhibitors.

Genotype-transcriptome and genotype-drug sensitivity correlations have not previously been described in Wilms tumor. Despite these different transcriptional effects, our analyses suggest that these divergent genomic features converge on the common downstream effect of CCND2 overexpression to drive Wilms tumor. CCND2 is known to be highly expressed in –80% of Wilms tumors^50,66, and other studies have associated CCND2 overexpression with SIX1/2 mutations⁵, MYC activation⁶⁷, Wnt signaling⁶⁸, and WT1 loss⁶⁹. Little is known about how miRNA pathway mutations drive Wilms tumor. Our study demonstrates that miRNA processing mutations are also associated with CCND2 overexpression, which drives proliferation. We also demonstrate that miRNA processing mutations restrain differentiation through de-repression of NPC marker genes. Because miRNAs commonly repress pluripotency genes during embryonic development¹², loss of miRNAs is associated with a self-renewing NPC transcriptome signature.

Compared to previous multi-platform sequencing reports in Wilms tumor, our approach provides certain advantages. First, prior sequencing efforts have not demonstrated an overall decrease in miRNA expression in Wilms tumors with miRNA processing mutations^5,6. Our use of a spike-in control to normalize miRNA read counts allowed us to detect global changes in miRNA abundance. Second, our previous in vitro studies predicted that heterozygous DROSHA missense mutations would produce “dominant-negative” global impairment of microRNA processing⁴. However, previous studies that measured miRNA levels in Wilms tumor by sequencing did not show such global effects^4–6. Similarly, this is the first report to document changes in pri-miRNA cleavage in human tumors, as this is the first to use whole-transcriptome sequencing instead of enrichment for polyadenylated transcripts.

Despite the comprehensive nature of our analysis, many Wilms tumors still lack an obvious driver mutation. In other words, in addition to the four genome-transcriptome signatures we describe, a fifth category of undefined tumors remains. These tumors may be driven by rarer mutations or combinations of mutations that cooperatively drive cancer formation. Furthermore, our analysis was not designed to identify changes in methylation, such as imprinting changes at the 11p15 locus, that are not thought to be sufficient for tumorigenesis⁷⁰. Future studies may identify driver mutations in noncoding regions or genes that are mutated less commonly, as we enter a new era of personalized medicine where clinical tumor sequencing becomes commonplace.

While there are currently no targeted therapies in use in Wilms tumor, our study highlights several potential avenues for future investigation. First, since Wilms tumors are driven by CCND2 signaling, they may be susceptible to inhibition of CDK4/6 or upstream mitogenic signaling pathways. Cancers in which D-type cyclins are overexpressed through amplification, translocation, or loss of the 3′UTR, where microRNAs bind, are particularly susceptible to CDK4/6 inhibitors⁷¹. Loss of microRNAs themselves may similarly sensitize these cells to CDK4/6 inhibition. CDK4/6 inhibitors are approved for some adult cancers and remain under investigation in pediatrics.

Second, while GPX4 inhibitors are not yet in clinical testing, this type of drug is under active development. While ferroptosis-related expression signatures may be enriched in Wilms tumor⁷², little is known about the therapeutic possibilities of ferroptosis induction in Wilms tumor. Our data suggest that Wilms tumors with microRNA processing mutations may be sensitized to pharmacological ferroptosis inducers due to their enhanced mesenchymal state. Since the mesenchymal blastema is thought to be the “cancer stem cell” of Wilms tumor, therapies that attack the mesenchyme could be effective in Wilms tumors driven by other mutations as well. In addition, since ferroptosis may be part of the mechanism by which some anticancer therapies work, ferroptosis inducers could enhance sensitivity to other treatments.

Implications:

This study reveals genotype-transcriptome relationships in Wilms tumor and points to ferroptosis as a potentially therapeutic vulnerability in one subset of Wilms tumor.