Deep sequencing of human papillomavirus positive loco-regionally advanced oropharyngeal squamous cell carcinomas reveals novel mutational signature

A custom panel of 395 cancer-associated genes (Additional file 1: Table S1) were targeted and sequenced in tumour-normal pairs. Genes were identified via cBioportal.​org and selected based on results from The Cancer Genome Atlas [18], Agrawal et al. [19], and Rubio-Perez et al. [20]. Median coverage of the targeted bases was 150X in the tumours and 200X in the matched normal tissue.

Deamination of cytosine bases to uracil is a known source of DNA damage occurring in FFPE blocks that may lead to cytosine-thymine conversion (artefacts) in PCR and/or sequencing reactions [21], that can be ameliorated with the use of an Uracil-N-Glycosylase [22]. Therefore, we used the commercial GeneRead DNA FFPE Kit (Qiagen, Hilden, Germany), which is based on specific removal of deaminated cytosine residues from FFPE DNA using the Uracil-N-Glycosylase enzyme. The kit was used with the manufacturer’s instructions, except for the use of double amount of proteinase K and deparaffination solution, and samples were left overnight for proteinase K digestion at 56 °C.

Genome libraries were prepared according to the Ovation Custom Target Enrichment System protocol (NuGen). Landing-probes were designed by NuGEN to target the chosen genes. Target enrichment and library preparation was done according to the manufacturer’s protocol (NuGEN Technologies, San Carlos, CA). Briefly, genomic DNA was fragmented, end-repaired and ligated with forward barcoded adaptors, followed by a bead purification step. The barcoded adaptor contained both an 8-nt sample-specific barcode and a 6-nt UMI. The latter was used to identify and remove duplicate reads. The reverse adaptor was annealed to the target regions and extended. The library amplification step comprised 25 PCR cycles followed by bead purification and sequencing. Sequencing was carried out on a NextSeq 500 (Illumina Inc., San Diego, CA) with single-end sequencing (150-bp) using the high-output kit v2.

Raw sequencing data (.bcl files) were demultiplexed into individual FastQ read files with Illumina’s bcl2fastq v2.16.0.10 (Illumina Inc., San Diego, CA) based on their unique index, with the R1 read containing the forward read and R2 containing the UMI. Each sequencing library was checked for quality with fastQC (version 0.11.5). A combination of cutadapt (version 1.11), BBDuk (version 35.82) and ERNE-filtering (version 2.1.1) was employed to remove adapter and linker sequences, remove trailing probe sequence on R1, and clip low-quality bases ends. Read alignment was carried out with the BWA algorithm (BWA version 0.7.10) with UCSC hg19 (GRCh37) as the reference. Deduplication was performed with an in-house script developed at IGA Technology (http://​igatechnology.​com), which makes use of the UMI to accurately remove PCR duplicates. The deduplicated reads were subsequently prepared for variant calling after realignment to correct for misalignments around indels and recalibrating base qualities based on known polymorphisms present in the human dbSNP (dbSNP137) with the GATK tool (v3.6–0-g89b7209). Variants were called with Mutect2 based on tumour-normal pairs with standard settings. Only mutations that passed Mutect2 filters were used in the downstream analysis. We performed a filtering step before annotation, where SNPs present in the European 1000 Genomes cohort [23] with an allele frequency above 1% were removed, as they are more likely to represent common population polymorphisms than somatic mutations. We also filtered SNPs that did not full-fill the following criteria: Number of mutant allele reads in tumour sample ≥ 5, total number of allele reads ≥20, quality score (QSS) > 20 and variant allele fraction > 0.1, in order better to exclude C > T variants associated with FFPE artefacts [24]. To account for false positive variants from Mutect2 samples with a SQSS > 20 was filtered, where SQSS is QSS/allelic depth in the tumour [25]. For annotation, we used Oncotator v. 1.8.1.0 [26]. For visualization of mutations we used maftools [27]. Somatic genes that were significantly mutated were identified using OncodriveFML 1.1 using a threshold of q < 0.1 with CADD v.1.3, a signature by cancer type and coding regions [28]. Mutational signatures were predicted in each sample using the COSMIC framework [29]. Substitution mutations across the whole genome were analysed in context of the flanking nucleotides (96 possible trinucleotide combinations). Identified signatures were compared with 30 other validated signatures, and the frequency of each signature per megabase was determined. The numbers of enriched mutational signatures were established by a best-fit rank based on decreasing cophenetic correlation coefficients. Apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like (APOBEC) enrichment scores were estimated as described by Burns et al. [30]. Briefly, enrichment of C > T mutations occurring within a tri-nucleotide DNA sequence (tCw) motif over all of the C > T mutations in a given sample was compared to background cytosine ratio and tCw’s occurring within 20 bp of mutated bases. Samples were classified as APOBEC-enriched based on an APOBEC enrichment score > 2 and/or false discovery rate (FDR) < 0.05. DNA copy number variations (CNV) were predicted by analysing the panel-sequencing data using CNVPanelizer, which compares tumour samples with a pool of normal tissue samples. The algorithm combines bootstrapping the reference set with the subsampling of amplicons associated with each of the target genes. This serves as a non-parametric estimation of the distribution of the gene-wise mean ratio between healthy reference samples and each tumour sample. All normal samples were used as a reference for calculating the sample-wise CNVs. Only genes that passed the noise and significance filter are reported as amplifications or deletions. To compare the difference in mutations between patients with high and low exposure to tobacco, we performed a fisher test on all genes between the two conditions to detect differentially mutated genes.

The predominant type of SNV was found to be C > T and T > C (Figs. 1 and 2). The DNA substitution mutations were most commonly transitions (Fig. 2). The mutational pattern of each sample was compared against all 30 validated mutational signatures from COSMIC [29]. This was assessed by the COSMIC algorithm that takes the sequence context of each mutation into account (Fig. 3). The tumours were dominated by COSMIC signature 5 or 12 (Figs. 2 and 3). Signature 5 has previously been observed to be enriched in HNSCC. The signature is depicted in most cancers and exhibits transcriptional strand bias for T > C substitutions at an ApTpN context [31]. Signature 12 is known to be present in a subset of liver cancers, and exhibits a strong transcriptional strand-bias for T > C substitutions [31]. It was also investigated whether an APOBEC signature was significantly enriched in our samples. The APOBEC signature was enriched in a quarter of the samples (23%) [30] (Additional file 2: Figure S1).

The OncodriveFML (q < 0.05) tool was employed to identify genes enriched in mutations. This algorithm identified 167 significantly mutated genes (Fig. 4). Of these APOB (35%), BIRC6 (32%), SPTBN1 (32%), FAT2 (32%), KMT2A (31%), FAT1 (30%), BPTF (30%), TRIO (30%), HERC2 (28%), and KALRN (28%) were among the most frequently mutated genes. The most prevalent genomic alterations were located in KALRN (28%), SPTBN1 (32%), KMT2A (31%), ZNRF3 (9%), BNC2 (12%), NOTCH2 (25%), FGFR2 (12%), SMAD2 (6%), and AR (13%).

We determined CNAs from our sequencing panel by employing CNVPanelizer, an algorithm specifically designed to estimate copy number frequencies using targeted massive parallel sequencing data. The most frequent gene copy gain was detected for ATF68 (n = 83; 72% of samples) and IL3RA (n = 81; 71% of samples). We found a high level of inter-patient CNA heterogeneity, but also a high intra-patient CN heterogeneity.

In the cohort, we identified the mutant-allele tumour heterogeneity (MATH) score. The median score was 20.8 (range 8–36) (Additional file 3: Figure S2).

Information on smoking was available for 70 patients with a median number of pack years of 15 (95% CL: 1.5–20 pack years). Thirty patients had never smoked or had less than 10 pack years. The genetic aberrations differed significantly between smokers and non-smokers. For smokers the genes FGFR2, EPHA2 were significantly more likely to be mutated. For non-smokers the genes IREB, BAZ2 and BRCA1 were significantly more often mutated. The mutational signature for those with low tobacco use was signature 5 and signature 12 in that order. The mutational signature for those with high tobacco use was in contrast 12 and 5. We did not observe any significant difference in MATH score between those with high and low tobacco use. There was also no significant increase in mean number of mutations per sample between these two groups.