Genetic analysis of sinonasal undifferentiated carcinoma discovers recurrent SWI/SNF alterations and a novel PGAP3-SRPK1 fusion gene

A single-institution retrospective case series informed by a prospectively maintained database of patients with SNUC was performed. The study was approved by the University of Michigan Institutional Review Board (HUM00080561). Patients with a history of sinonasal undifferentiated carcinoma treated at the University of Michigan were included in the clinical dataset (n = 46). Pathology descriptions for the cohort are listed in Supplemental Table 1a. Inclusion criteria for genomic, copy number and transcriptome analysis is as follows: 1) Patients with sinonasal undifferentiated carcinoma as confirmed by our board-certified pathologist (J.B.M.); 2) Blocks maintained in the University of Michigan pathology archive; 3) Sufficient DNA or RNA yield for next generation sequencing. Additionally, a prospective patient was consented to our University of Michigan IRB-approved MiOTOseq precision medicine program (HUM00085888) as described [20]. In total, there were 5 patients who met inclusion criteria for analysis, and demographics are shown in Supplemental Table 1b.

The patient derived SNUC cell line, MDA8788–6, was generously provided by MD Anderson. Generation of this cell line was previously described by Takahashi et.al [21]. Cells were cultured in a humidified incubator at 37 °C with 5% (vol/vol) CO2 in DMEM with 10% FBS, 1X Pen/Strep, 1X NEAA. Cells were genotyped to confirm the STR profile of the cell line (Supplemental Table 2) as previously described [22].

DNA was isolated from formalin fixed, paraffin-embedded (FFPE) samples following the manufacturer’s protocol for AllPrep DNA/RNA FFPE kit (Qiagen, Hilden, Germany) as previously described [23, 24]. Tumor and adjacent normal regions were identified on H&E stained slides and aligned to tissue paraffin blocks. An 18-gauge sterile needle was used to core 2–4 samples from each region. Deparaffinization was performed using the xylene/ethanol method with the only modification being that samples were digested using proteinase K at 56 °C for 20–24 min. DNA isolation was then completed using the Allprep Isolation kit (Qiagen, Hilden, DE) following manufacturer protocol. Each sample was analyzed using a Nanodrop spectrophotometer for purity (260:230 and 260:280 ratios) and concentration was determined using 1uL of sample with the Qubit 2.0 Fluorometer and measured with a bioanalyzer as described [25]. DNA extraction for MDA8788–6 cell line was performed using Wizard® Genomic DNA Purification Kit (Promega, Madison, WI).

Genomic DNA from each tumor and adjacent normal specimen was submitted for sequencing to the University of Michigan’s DNA sequencing core for exome sequencing using both the DNA TruSeq Exome Library Preparation kit (Illumina, Catalogue number FC-150-100x; SNUC2, SNUC5, SNUC8, SNUC10) and the Roche NimbleGen V3 capture kit (SNUC1). DNA from the MDA 8788–6 cell line was sequenced as described [26]. Libraries were prepared according to the manufacturer’s instructions. Libraries were then paired end sequenced to 125 nucleotides as part of pool with an average of 4 samples per lane on an Illumina HiSEQ4000 yielding an average depth of greater than 90x per sample.

Quality of the sequencing reads was assessed using FastQC v.0.11.5. Because the reads had adapter contamination as well as a high k-mer content at the start of the reads, trim galore v0.4.4 was used to remove adapters and trim reads. Reads were aligned to the hg19 reference genome using BWA v0.7.1. Mapping was followed by marking duplicates using PicardTools v1.79. Base quality score recalibration was done using GATK v3.6 and this was the last step in preparing the reads for variant calling. Samtools v1.2 was used to create pileup files for each tumor-normal pair. Varscan v2.4.1 was used to call variants from these mpileup files using the somatic mode of the variant caller. Goldex Helix Varseq v1.4.6 was used to annotate variants. All variants in the introns and intergenic regions were filtered out. Variants with more than 5 reads supporting the alternate allele in the tumor samples were considered as true positives.

Aberration Detection in Tumor Exome (ADTEx) v.2.0 was used to make copy number estimation calls from the pre-processed tumor-normal BAM files which were also used for variant calling. A state from 0 to 4 was assigned by the software based on its estimated copy number. State 0 corresponds to a homozygous deletion, 1 corresponds to a heterozygous deletion. A normal copy number is denoted by state 2. States 3 and 4 represent a gain and amplification respectively.

MSIsensor was used to detect somatic MSI loci from the tumor-normal sample pairs as described previously [27]. The software assigns a status to each sample pair based on an instability score calculated based on a threshold of more or less than 3.5% of called microsatellites having alterations. We present this score as well as the overall percentage of microsatellite alterations for each tumor-normal pair.

Excess DNA from above was used to validate mutation calls in novel genes. Primers were designed using MITprimer3 to amplify a small region surrounding the nominated single nucleotide variants (SNVs) as described in Supplemental Table 3. Polymerase chain reactions (PCRs) were optimized for each primer pair on cell line genomic DNA and then used to amplify the regions from tumor and adjacent normal DNA using Platinum Taq DNA High Fidelity polymerase (ThermoFisher, Waltham, MA). PCRs products were then visualized on an eGel (Invitrogen, Waltham, MA) and purified using the a PCR purification kit (Qiagen, Valenica, CA) as described [28] and submitted for Sanger sequencing at the University of Michigan’s DNA sequencing core. Results were visualized using the LaserGene software suite.

Total RNA from the MDA8788–6 sample underwent standard QC and was submitted for RNA sequencing to the University’s DNA sequencing core as previously described [26, 29]. Briefly, the Illumina Stranded RNAseq kit was used and libraries were sequenced on an Illumina HiSEQ4000 using 75 nt paired end approach. Quality of the RNA sequencing reads was determined using FastQC v0.11.5 and we did not identify any quality issues. We then used a two-step alignment protocol of Star v2.5.3a to map the reads and genome index files were first generated using the reference human genome and annotated transcriptome files. In the second step, we then used the index files to guide read mapping. Samtools v1.9 and Picard v2.4.1 were used to retain only uniquely mapped reads and FPKM was computed using Cufflinks v2.2.1 with default parameters, with the exception of modifying “--max-bundle-frags” to 100,000,000. This modification was made to avoid raising of the HIDATA flag at loci that have more fragments than the pre-set threshold for every locus.

FusionCatcher (v1.00) is a software package designed to look for gene fusions, translocations and rearrangement events using paired end RNA-Seq data and was used to identify novel gene fusions in the MDA8788–6 cell line.

High molecular weight DNA was isolated from the SNUC cell line by lysing 1.5 million cells overnight at 37° with lysis buffer (10 mM Tris-HCl, 400 mM NaCl, 2 mM EDTA), 10% SDS, and a proteinase K solution (1 mg/mL Proteinase K, 1% SDS, 2 mM EDTA). Following overnight lysis, DNA was salted out of the solution with 5 M NaCl for 1 h at 4° and precipitated with ice cold ethanol for 5 h at − 20 °C. High molecular weight DNA was eluted in TE buffer; the quality and integrity of the DNA was assessed using the Tapestation Genomic DNA ScreenTape kit (Agilent). The DNA was submitted to the University’s DNA sequencing core for 10x based linked read library generation and sequencing on an Illumina NovaSeq6000 with 300 nt paired end run. Samples were de-multiplexed and FastQ files with matched index files were generated using Long Ranger Version 2.2.2. Data was visualized using Loupe software package, Version2.1.1 (2.4).

All siRNA including ON-TARGETplus Non-targeting Control siRNA, ON-TARGETplus GAPDH Control siRNA, and a custom siRNA targeting the PGAP3-SRPK1 fusion site were purchased through Dharmacon (Lafayette, CO). Each siRNA was reconstituted at a concentration of 1 nmol/50 uL in 1X siRNA buffer (DHarmacon, Lafayette, CO). MDA8788–6 cells were plated at a concentration of 250,000 cell per well in 3 mL growth media. The following day all media was removed and cells were starved in 1 mL of serum DMEM for 3 h. Each siRNA was prepared by adding 400uL of OPTI-MEM (Gibco, Waltham, Massachusetts) with 24uL of siRNA and left to equilibrate for 5 min. Separately, 24uL of oligofectamine (Invitrogen, Carlsbad, CA) was added to 96uL of OPTI-MEM. After 5 minutes the two mixtures are added together and allowed to equilibrate at room temperature for 20 min. Cell were then treated with 250uL of siRNA mixture containing buffer only, Non-targeting siRNA, PGAP3-SRPK1 fusion siRNA, or GAPDH siRNA. After 3 h 2.5 mL of growth medium was added to each well. The following day cells were harvested in 700uL of QIAzol Lysis Reagent (Qiagen, Valencia, CA) and proceeded directly to RNA extraction or stored at minus 80 °C for future extraction. RNA extraction was performed using RNeasy Mini Kit (Qiagen, Valenica, CA) per manufacturer’s instructions. RNA sequencing of the fusion knockdown was also performed as above. Briefly, extracted total RNA from MDA8788–6 NT siRNA and PGAP3-SRPK1 fusion siRNA were submitted to the University’s DNA sequencing core and processed as above (Illumina Stranded RNAseq kit was used and libraries were sequenced on an Illumina NovaSEQ6000 using 300 cycle paired end approach).

Confirmation of successful siRNA knockdown and validation of RNAseq findings were performed with qPCR. Following RNA extraction, cDNA synthesis was performed using SuperScript™ III First-Strand Synthesis System (Invitrogen, Carlsbad, CA) and qPCR was performed using QuantiTect SYBR Green PCR Kit (Qiagen, Valencia, CA) and run on QuantStudio5 (Applied Biosystems, Foster City CA). Targets included SRPK1, PGAP3-SPRK1 fusion, GAPDH, HSDL2, CCND1, FOXO4, Beta-Actin, HRPT, and RPL-19; primer sequences are listed in Supplemental Table 4. Analysis was performed using the ^2ΔΔ-Ct method [30].

Forty-six patients were included in the survival analysis. The median age at diagnosis 53 years with a range from 19 to 87 years. Median follow up time was 28 months (range <  1 month – 23 years). Two patients (4.3%) were treated with surgery alone, 21 patients (44%) with surgery in addition to adjuvant radiation, chemotherapy, or both (CRT) and 23 patients (49%) with CRT alone. Twenty-five patients (53%) had persistent or recurrent disease after treatment. The median time to recurrence was 2.8 months. Of these 64% were locoregional recurrence or persistence while 32% were distant failures. Five-year overall survival was 42% (95% CI 27–56%) and 5-year disease specific survival was 46% (95% CI 30–61%) as seen in Fig. 1a. There was no significant difference in 5-year DSS when stratifying by use of surgery +/− CRT and CRT alone (50% [95% CI 27–69%] vs. 49% [95% CI 25–68%], p = 0.85), Fig. 1b. Multivariate analysis was performed using backward Wald cox regression. Variables included in the model were T-stage, nodal disease, tobacco use and treatment type (surgery +−/ CRT vs. CRT alone). Only tobacco use was found to be significantly associated with decreased survival (HR 5.1 [95% CI 1.7–15], p = 0.004).

We were able to isolate high quality DNA that met our quality control standards for sequencing from 5 retrospective SNUC and matched normal samples that were advanced for full exome sequencing. Within our 5 sample cohort, we had 2 patients that died within 2 years of diagnosis and 3 patients that survived for more than 5 years after diagnosis (Table 1). Using these tumor and adjacent normal DNA samples, we sequenced exomes to an average depth of 256,352,493 yielding an average of 218,566,135 uniquely mapped reads in each sample, for an average coverage of >100X per tumor (Supplemental Table 5).

Using this data, we first generated copy number calls for broad genomic regions and assessed the global view of the copy number variation in Circos plots (Fig. 2a, Supplemental Tables 6A-E). These plots demonstrated that 3/5 tumors showed a much higher level of copy number variation as compared to the other samples with frequent high level amplifications in chromosomes 4, 17, 19 and 22. Genes with copy gain or amplifications in multiple tumors included: CD300C which plays a role in innate immunity and antigen presentation via MHC class I and is a negative regulator of CD4 and CD8 T cells [31], JAK3, E2F4 and GLI1, which have canonical and non-cannonical roles in tumor cell proliferation respectively [32, 33], MDM2 which targets p53 for degradation [34], and MLL2 which may play a role in epithelial-mesenchymal transition (EMT) [35]. We also identified copy gains in ERBB2 in 3/5 samples (Fig. 2b) suggesting a potential role of ERBB2 function in a subset of SNUCs, which has been previously noted [18]. Moreover, we also identified copy number losses in SMARC family genes (SMARCA1, SMARCA2, SMARCA5, SMARCB1, SMARCC1 and SMARCE1) in numerous samples consistent with previous reports of the recurrence of alterations to this gene family in SNUC [16]. Focal amplifications were identified in WEE1, FGFR3 and MAPK15 while focal copy losses were identified in FAT1 and SMARCA2 (Fig. 2c).

Analysis of SNV data revealed an average of 23.6 (range of 5–63) high confidence non-synonymous somatic alterations per tumor (Supplemental Figure 1). Oncoplots for recurrently altered as well as targetable or cancer-relevant genes are shown in Fig. 3a and an extended list of the mutations with functional annotations and pathogenicity scores is shown in (Supplementary Tables 7, 8, 9). From this analysis, we identified a mutation in IDH2 (p.Arg172Gly), which is a gene previously associated with SNUC [15, 17], as well as new mutation disrupting SMARCAL1 (p.Thr742Met; Fig. 3b, c) suggesting alternative pathways may be implicated in tumorigenesis in these samples. Further, we identified potentially targetable alterations including an ALK p.Gly872Ser (Fig. 3d). We also identified two different ABC transporter genes, ABCA10 (p.Glu507Ter) and ABCB7 (p.Glu342Gln), suggesting a general pathway for drug resistance that may explain why some tumors respond poorly to chemotherapy.

Finally, we characterized microsatellite instability in each of the tumors using the MSIsensor software package, which assigns an instability score to each tumor/normal pair. These data revealed an average of 16.4 somatic sites and a median % of 4.99 (range 0.72–6.95). Two of the SNUC samples were found to be microsatellite stable while three were found to be unstable (Supplemental Table 10).

Our colleagues recently derived the first SNUC cell line, MDA8788–6, and we performed exome sequencing on this cell line model with subsequent SNV annotation using our previously established informatics pipelines for head and neck cell lines without available matched normal DNA. This sequencing yielded 225,772,526 reads, of which 99.7% uniquely mapped to the genome yielding an average coverage of >100X (Supplemental Table 11). As no matched normal was available, joint calling was completed with our UM-SCC cell lines as described [26, 29], and filtering to retain only heterozygous calls and remove any SNV previously reported in dbSNP, yielded 563 potential single nucleotide variants, of which 182 were categorized as missense, frameshift or stopgains (Supplemental Table 12). Similar filtering for insertion/deletion calls yielded 779 potential INDELs, of which 39 were categorized as frameshift INDELs (Supplemental Table 13). Among these alterations, we identified KMT2B S421P and NOTCH1 T2483 missense mutations, which may have an important role in pathogenesis of this tumor model given the previously established role of these genes in tumorigenesis.

Next, we performed paired end RNA sequencing to discover potential gene fusions in the cell line using the FusionCatcher algorithm. This analysis called 107 potential gene fusions in the cell line to a variety of known oncogenes and previously described fusion genes from other cancers (Supplemental Table 14). We focused validation studies on inframe gene fusions, with at least 10 supporting reads that spanned the fusion junction. Importantly, we were able to validate the presence of a novel fusion between PGAP3 exon 2 and SRPK1 exon 13 (Fig. 4a). The resulting fusion gene creates an in-frame fusion gene predicted to encode a 244 amino acid, with predicted isoelectric point (pI) of 5.87 and molecular weight of 27.6 kDa protein (Supplemental Figure 2) [36, 37]. The resulting fusion protein is predicted to retain the SRPK1 protein kinase domain, which normally regulates constitutive and alternative pre-mRNA splicing machinery through phosphorylation of SR proteins [38].

To then test the hypothesis that large scale structural rearrangements drove fusion formation, we performed 10X linked read genome sequencing on high molecular weight DNA isolated from the MDA8788–6 cell line. We obtained 848,842,686 total reads and averaged 36.9x coverage, with 96.1% of the DNA molecules of > 20 kb in length and 58% greater than 100 kb (Supplemental Table 15). Long ranger pipeline analysis phased 99.2% of SNPs resulting in a maximum phase block of 47.9 MB (Fig. 4b). The analysis identified 888 large structural variant calls in the genome. Contrary to our hypothesis, there were no structural rearrangements around PGAP3 or SRPK1 identified (Fig. 4c-e) suggesting that the fusion is formed at the RNA level, possibly by trans-splicing events. However, we did identify structural alterations involving ZNF546 and AXL (Chr19:40,490,000 and Chr19:41,760,000, 1.27 MB duplication, quality score 1000) as well as CREB3L2 and BRAF (Chr7:137,630,000 and Chr7:140,490,000, 2.86 MB duplication) suggesting a potential role for these oncogenes in pathogenesis of this SNUC model (Supplemental Figure 3).

Finally, to test for potential functional roles of the PGAP3-SRPK1 fusion, we performed siRNA-mediated knockdown of the fusion transcript and submitted RNA for complete transcriptome sequencing. This data demonstrated that knockdown of PGAP3-SRPK1 fusion results in significantly decreased HSDL2, NAGK, and CCND1 RNA expression and a modest increase in LINC01006 and FOXO4 suggesting PGAP3-SRPK-1 may play a role in regulation of these genes (Fig. 5a). In fact, we identified 122 genes that were > 2-fold upregulated and 112 genes > 2-fold downregulated in the knockdown relative to control. Further, gene set enrichment analysis of the differentially expressed rank list found enrichment of KOBAYASHI_EGFR_SIGNALING_24HR_DN, HALLMARK_E2F_TARGETS, and HALLMARK_MYC_TARGETS_V1 pathways (Fig. 5b, Supplemental Table 16). Collectively, this RNA sequencing data analysis suggests that the gene fusion has a pivotal role in the hallmark signaling pathways.