Significance and limitations of the use of next-generation sequencing technologies for detecting mutational signatures

Abstract

Next generation sequencing technologies (NGS) have been critical in characterizing the genomic landscape and untangling the genetic heterogeneity of human cancer. Since its advent, NGS has played a pivotal role in identifying the patterns of somatic mutations imprinted on cancer genomes and in deciphering the signatures of the mutational processes that have generated these patterns. Mutational signatures serve as phenotypic molecular footprints of exposures to environmental factors as well as deficiency and infidelity of DNA replication and repair pathways. Since the first roadmap of mutational signatures in human cancer was generated from whole-genome and whole-exome sequencing data, there has been a growing interest to extract mutational signatures from other NGS technologies such as targeted panel sequencing, RNA sequencing, single-cell sequencing, duplex sequencing, reduced representation sequencing, and long-read sequencing. Many of these technologies have their inherent sequencing biases and produce technical artifacts that can confound the extraction of reliable and interpretable mutational signatures. In this review, we highlight the relevance, limitations, and prospects of using different NGS technologies for examining mutational patterns and for deciphering mutational signatures.

INTRODUCTION

Technological and engineering advances in DNA sequencing approaches have revolutionized the field of cancer biology. The past two decades were marked by a rapid paradigm shift from the use of Sanger sequencing by capillary electrophoresis for interrogating particular exons in a handful of known cancer genes to a next-generation of sequencing approaches (NGS) allowing massively parallel high-throughput sequencing of complete cancer genomes [1]. Analyses based on NGS data have facilitated the characterization of cancer as an evolutionary disease governed by the principles of Darwinian evolution and fueled by the acquisition of genomic alterations [2]. With the high-resolution molecular lens provided by NGS, we can now elucidate and time the accumulation of these genomic alterations throughout the lineage of a cancer cell. The process of mutagenesis, driven by both the activity of endogenous and exogenous DNA damaging agents as well as the infidelity or defects of intrinsic cellular pathways (e.g., pathways responsible for DNA repair or replication), starts as early as the first division of the fertilized egg and continues to operate after a normal cell undergoes a neoplastic transformation [2]. Thus, the complete catalog of genomic alterations found in a cancer genome can serve as an archeological footprint of the historical processes that have been operative throughout the lineage of the cancer.

For much of the last sixty years, the field of cancer biology has been dominated by the somatic mutation theory which examines cancer as a disease of the genome driven by the successive accumulation of somatic mutations [3]. This has led to a number of initiatives utilizing different technologies for mapping the genomic events that are essential and sufficient for tumorigenesis. Initially, cytogenetic analyses were conducted to identify oncogenic large-scale chromosomal aberrations or defects in the nuclei of cancer cells. The development of Sanger sequencing in the 1970s [4] propelled an avalanche of cancer research focused on vigorously categorizing both somatic mutations and commonly mutated genes in different cancer types [5–7]. With the advent of next-generation sequencing in the mid-2000s, it became possible to sequence large numbers of cancer genomes with high accuracy at a relatively low cost [8]. This propelled the coordinated efforts of the International Cancer Genome Consortium (ICGC) and The Cancer Genome Atlas (TCGA) [9,10] to generate large-scale datasets that characterize the genomic landscapes of cancers in thousands of patients with more than 40 tumor types. These datasets provided sufficient power to unbiasedly identify and characterize the mutational processes operative across human cancer. In 2013, a novel computational framework, based on nonnegative matrix factorization, was developed to decipher distinct mutational signatures across cancer genomes [11]. This framework allowed moving from mutational patterns, which are directly observable from the sequencing data, towards mutational signatures, which are not directly observable from the data but allow explaining the observed patterns of mutations. For example, if one sequences a sample and observes 2,500 C:G>T:A mutations and 7,500 C:G>A:T mutations, the mutational pattern of the sample will be 10,000 single base substitutions with 25% C:G>T:A and 75% C:G>A:T mutations. The pattern of mutations in this example sample may have been generated by the signatures of four distinct mutational process: (i) methylation followed by deamination generating C:G>T:A mutations at NpCpG context; (ii) enzymatic deamination driven by the APOBEC enzymes leading to C:G>T:A substitutions at TpCpN context; (iii) activity of reactive oxygen species causing C:G>A:T mutations; and (iv) exposure to tobacco smoking imprinting C:G>A:T mutations. Under the developed computational framework, the single base substitutions in a cancer genome were examined as a 96-dimmensional vector, where each dimension reflected one of the 6 possible types of base substitutions (C:G>A:T, C:G>G:C, C:G>T:A, T:A>A:T, T:A>C:G, and T:A>G:C) and their immediate 5’ and 3’ flanking bases. An initial application of this framework identified 21 distinct mutational signatures across 7,042 primary cancers [12]. With subsequent improvements in the algorithms for signature extraction and the increased availability of whole-genome sequenced cancers, we now have a reference set of 49 distinct single base substitution (SBS), 4 clustered base substitution, 17 insertions and deletions (ID), and 11 doublet-base substitution (DBS) signatures across more than 20,000 cancer genomes [13,14]. Thus far, more than 20 structural variation (SV) signatures have also been identified [15,16]. Additional SBS, ID, DBS, and SV signatures, not part of the reference signatures, have been identified in numerous other independent studies [17,18].

Since the initial roadmap of mutational signatures was published in 2013, a lot of progress has been made in identifying the underlying factors driving mutational patterns observed in cancer genomes, their broad-term implications in exploring cancer prevention strategies and utilizing mutational signatures for clinical decision making. Examples of signatures driven by exogenous factors include SBS4 and SBS7 that result from tobacco smoking or exposure to ultraviolet radiation from sunlight, respectively. SBS4 is mainly comprised of C:G > A:T transversions, it is highly enriched in tobacco smokers [19] and shares a similar mutational profile to the one imprinted by the tobacco carcinogen benzo[a]pyrene in experimental systems [19]. SBS7 is characterized by C:G>T:A substitutions at dipyrimidines, it is almost exclusively found in skin cancers [12,13] and its pattern has been reproduced in a plethora of experimental systems [20]. The majority of mutational signatures attributed to endogenous factors are due to infidelity or failure of DNA repair/replication pathways. For example, SBS3 is a flat mutational signature enriched in homologous recombination deficient (HRD) breast and ovarian cancer patients [12,21]. HRD deficient cell lines have been shown to accumulate SBS3 [22] and cell lines harboring SBS3 were shown to be sensitive to Poly (ADP-ribose) polymerase (PARP) or platinum therapy [23]. Machine learning-based prediction tools have been developed to incorporate SBS3 and other HRD signatures as clinical biomarkers in order to predict patient response to PARP or platinum therapies [23–25].

In the process of discerning the etiology and the clinical utility of mutational signatures, a number of mutation calling tools [26–29] and computational frameworks for extracting mutational signatures [11,30–34] have been applied to extract mutational signatures across different tissue types and sequencing platforms. Degasperi et al. (2020) explored the limitations of current signature assignment approaches and proposed the use of tissue-specific instead of pan-cancer derived reference signatures to assign signatures to samples from different tissue types [18]. On one hand, the use of tissue-specific reference signatures will eliminate the possibility of overfitting and misassigning biologically irrelevant signatures to samples [18]. On the other hand, using a prior predefined set of tissue-specific reference signatures introduces a bias in the signature assignment process where the chances of discovering new signatures or identifying known signatures decrease in cancer types where these signatures were not previously found. Additionally, differences in variant discovery, signature extraction approaches, and signature assignment methods have introduced complications in reliably standardizing the analysis of mutational signatures. To address these limitations, prior research has systematically compared different signature extraction methods [35] and proposed a standard workflow for de novo extraction of mutational signatures [36].

Nonetheless, several parameters independent of the signature extraction and assignment process can impact the specificity and sensitivity for detecting mutational signature. These parameters include: (i) number of cancer genomes; (ii) degree of similarity between different operative mutational signatures; (iii) number of mutational signatures found in a cancer genome; (iv) strength/exposure of mutational signatures; (v) overall tumor mutational burden (TMB) of individual samples; (vi) the number of mutation channels used to examine the operative signatures [11,37]. The number of required cancer genomes to identify all active signatures increases exponentially with the total number of operative signatures [11]. The strength of a signature’s exposure can impact the discovery of other signatures present at low levels (less than 5% contribution) across a set of cancer genomes [11]. A high tumor mutational burden allows for reliable signature extraction; thus, having fewer cancer genomes with a high TMB is more critical than many cancer genomes with low TMB [11]. The number of mutation types/channels can aid in the discovery of additional signatures. However, with more channels, fewer mutations exist per mutation type, limiting the ability to robustly identify individual mutational signatures [11].

The use of different sequencing technologies can drive differences in the mutational patterns observed in cancer samples as well as the mutational signatures extracted from these patterns [13]. Different NGS technologies have inherent sequencing biases and technical artifacts specific to a library preparation protocol or to a sequencing platform that can impact the reliable discovery of signatures operative in a cancer genome [13]. Due to the lack of available literature on the use of different sequencing technologies for analysis of mutational patterns and mutational signatures, this review aims to systematically explore and address the applicability, significance, and limitations of different sequencing technologies for identifying mutational profiles and extracting mutational signatures. This review will not go over the etiologies or the underlying mechanisms driving the distinct mutational signatures as these have been extensively discussed in a number of recently published reviews [38–41].

Sanger Sequencing by Capillary Electrophoresis

Traditional Sanger sequencing, developed by Frederick Sanger in the 1970’s, was labor intensive and required both radioactively labeled DNA primers and slab gel electrophoresis [4]. As part of the Human Genome Project, Sanger sequencing was automated and based on capillary electrophoresis of individual fluorescently labelled sequencing reaction products [42]. This capillary sequencing approach was widely used for the rapid examination of targeted region(s) of the genome and played a pivotal role in aiding our understanding of cancer as a genetic disease. Capillary sequencing had an instrumental role in establishing the first set of mutational patterns operative across different cancer types using mutations found in the most commonly mutated cancer gene, TP53 [43,44]. Many studies grouped mutations found in some or all exons of TP53 across multiple samples to identify cancer-type specific mutational profiles. These studies reported heterogeneous and distinct TP53 mutational profiles reflective of the distinct mutational processes operative in different cancer types, including: an abundance of CC:GG>TT:AA mutations associated with sunlight exposure in skin cancers, an enrichment of C:G>A:T mutations induced by aflatoxins in hepatocellular cancers, and C:G>A:T mutations in tobacco smokers with lung cancer [43,44].

Single gene-specific mutational analyses have played a crucial role in identifying the distinct endogenous and exogenous processes operative across different cancer types; however, such studies require aggregating mutations from multiple samples to generate cancer-type specific mutational profiles. When an individual sample only contributes a few mutations to the overall mutational profile, it becomes hard to untackle and identify the different processes that operate at a single sample resolution [45]. Additionally, mutations comprising the mutational profile can also be affected by the selection pressures (both positive and negative) in the coding regions of genes. Thus, it becomes hard to determine whether the mutations in the mutational profile were preferentially enriched due to selection or generated due to the operative mutational processes [46]. Overall, the low mutational burden and the potential bias due to selection make Sanger sequencing suitable for identifying combined mutational patterns of strong mutagens but inadequate for more subtle mutational patterns or for extracting mutational signatures. These limitations are overcome by high-throughput sequencing technologies that enable the profiling of larger parts of the genome across many samples.

Next generation sequencing technologies (NGS) have been instrumental in cancer genomics and in the integration of precision oncology in the clinic. NGS’ high throughput capability, accuracy, and low sequencing costs enable the generation of large regions of the genome across many samples [1]. NGS technologies provide the capabilities for digital readouts of different molecular features altered in cancer genomes, including DNA and RNA at bulk and single-cell resolution.

Whole-genome and Whole-exome Sequencing

Whole-genome sequencing (WGS) and whole-exome sequencing (WES) technologies have been crucial for the identification of mutational patterns and the extraction of mutational signatures from cancer genomes. In 2013, the repertoire of mutational signatures was obtained from 6,535 whole-exome and 507 whole-genome sequenced cancers providing a molecular lens to decipher the different types of processes active in these cancers [12]. Through mutational signatures extracted primarily using exomes, we were able to recognize a number of endogenous processes, including: (i) widespread aberrant activity of the APOBEC3 family of deaminases; (ii) infidelity of certain DNA polymerases; and (iii) defects in DNA repair and replication pathways. These results aided our understanding of the inherent preferences of DNA repair pathways in fixing specific DNA lesions with underlying transcription and/or replication strand asymmetries [12,47,48]. The repertoire of mutational signatures also allowed for identifying the activity of exogenous DNA damaging agents such as: (i) polycyclic hydrocarbons in tobacco smoke; (ii) ultraviolet radiation in sunlight; and (iii) mutational patterns associated with treatments such as temozolomide in malignant melanomas and glioblastoma multiforme. More importantly, these signatures highlighted the inter-tissue variation in the mutational signature landscape; certain signatures were only prevalent in specific cancer types such as SBS16 in hepatocellular carcinomas and SBS10 associated with POLE proofreading deficiency in colorectal and endometrial cancers [12].

While the mutational signature analysis in 2013 was instrumental in identifying the underlying patterns of genomic alterations prevalent in cancer genomes, it had its limitations. This study could not provide more precise resolution of the features defining each signature and was limited in its ability to separate both very similar signatures (i.e., cosine similarity >0.90 between the 96 mutational profiles) and correlated signatures (i.e., signatures that operate commonly with the same proportions within the same samples such as APOBEC3 signatures SBS2 and SBS13). Using the complete mutation catalogue of 2,778 whole-genome sequenced cancers from the Pan-Cancer Analysis of Whole-Genomes (PCAWG) project in addition to other whole-genome and exome samples, in 2020, the most comprehensive pan-cancer mutational signature analysis provided many folds greater number of mutations that resulted in an increased power to extract rare and known signatures in cancer types that were not previously reported. This most recent pan-cancer analysis also allowed for the extraction of signatures associated with other mutation classes such as small insertions and deletions and doublet-base substitutions, which, in almost all cases, could not be deciphered from whole-exome sequencing data [13,16].

Differences in the resolution and the sensitivity of mutational signatures extracted using WGS and WES can be explained by the technology-specific differences that impact variant discovery. Systematic differences between the number and type of mutations identified in exons using both WGS and WES have been reported. Belkadi et al. (2015) reported a higher false positive rate for WES (78%) compared to WGS (17%) on mutations called from exons using WGS and WES for 6 unrelated individuals [49]. They also reported that 3% of the high-quality coding mutations found in WGS samples were missed by WES likely due to sequencing biases inherent to WES technology. More recently, Bailey et al. (2020) conducted a retrospective evaluation of mutations identified in exonic regions of 746 samples that have undergone both whole-exome and whole-genome sequencing. They reported 79% concordance between coding single base substitutions and 57% concordance between coding indels identified in WGS and WES mutation calls [50]. In general, sequencing errors such as PCR bias, variability in genome coverage in both high and low GC rich regions, and low genotype quality are more prevalent in WES than in WGS data [49–51]. Technical methods employed for WES and WGS that drive these reported technology-specific biases have been extensively discussed in several prior publications and will not be covered in this review [52–54].

The systematic limitations of WES and WGS leads to the question whether mutational signatures extracted by WES can recapitulate the mutational signatures extracted from exonic mutations identified in WGS. Bailey et al. (2020) reported a 90% Pearson correlation between the dominant set of mutational signatures identified in 746 samples across both exomes and down-sampled whole-genomes, indicating that the analysis of mutational signatures is robust between whole-exome and whole-genome sequenced samples [50]. In general, it may also be worth exploring the overall concordance between the patterns of the 96 mutational profiles extracted from the whole-genome sequenced PCAWG samples and whole-exome downsampling of these PCAWG samples. Comparing the mutational profiles of samples at whole-genome and exome resolution revealed that the overall similarity between the mutational profiles of the two sequencing technologies is greatly dependent on the tumor mutation burden (TMB) of each sample (Fig. 1A). Samples that belong to the class of tumors with high TMB such as skin-melanoma, lung, and colorectal cancers (Fig. 1A) have generally similar mutational profiles at both the whole-genome and whole-exome resolution. Note that a cosine similarity threshold of at least 0.85 is used to consider two profiles similar as this threshold corresponds to a p-value below 0.001 [55]. In general, tumor types with high TMB have a higher fraction of samples with consistent mutational profiles generated based on WES or WGS data (Fig. 1A).

Figure 1. Cosine similarities between the 96 mutational profiles of PCAWG samples generated from whole-genome sequencing and PCAWG samples down-sampled to whole-exomes.

A) Cosine similarity between the 96 mutational profiles of PCAWG samples at whole-genome (WGS) and exome (down-sampled WGS) resolution, split by tumor type. Top panel: somatic mutations per Megabase (Mb) (y-axis) and cosine similarities between the mutational profile of WGS and down-sampled WGS for each sample in a tumor type. Samples are colored by their cosine similarity: cosine ≥ 0.85 (orange) and cosine < 0.85 (dark blue). Bottom panel: The proportion of samples in each tumor type with mutational profiles (cosine ≥ 0.85) at whole-genome and whole-exome resolution.

B) Average normalized 96 mutational profiles of all 86 samples of the CNS-PiloAstro cancer at the whole-genome and exome resolution with a cosine similarity of 0.89.

C) Cosine similarity between the average normalized 96 mutational profiles at whole genome and exome resolution across different tumor types.

While the mutational profiles of individual samples vastly differ between WES and WGS data for cancer types with low TMB, the average mutational profiles obtained using the two technologies for a given cancer type is, in most cases, very similar. For example, while the patterns of individual CNS-PiloAstro cancers deferred at exome and genome resolution (Fig. 1A), the average normalized mutational profile of all 86 samples of the CNS-PiloAstro cancer type at the exome resolution closely resembles its average mutational profile at the whole-genome resolution (cosine similarity:0.89; Fig. 1B). The same observation holds for most cancer types with reasonable sample sizes that have very similar average mutational profiles at both exome and the whole-genome resolution (Fig. 1C). The high similarity between the average mutational profiles of WGS and WES data can be attributed to the use of a linear model for signature extraction that typically fails to accurately describe signatures in individual samples with low TMB but does a good job identifying these signatures over a large collection of samples.

Comparing the prevalence of mutational signatures of single base substitutions across samples profiled using WGS or WES approaches in the most recent pan-cancer analysis reveals clear technology-specific differences (Fig. 2). While many of the mutational signatures are found in the same proportions in most cancer types, some are clearly enriched in WES and WGS data. For example, signature SBS45, an artifactual signature attributed primarily to exome library preparation [56] was found to be much more prominent in whole-exome sequencing data. Similarly, signature SBS22, attributed to aristolochic acid exposure, was also found to be more prevalent in WES samples. This is likely due to bias in sample acquisition as most cancers with this type of environmental exposure come from specific regions of the world (viz., South East Asia) and these samples have been predominantly profiled using WES [57]. In contrast, flat mutational signatures (e.g., SBS3 - attributed to defective homologous recombination [21], SBS5 - attributed to clock-like ageing processes [12], and SBS40 with an unknown etiology [13]) and signatures found at lower mutational burden (e.g., SBS18 - attributed to reactive oxygen species [12]) can be detected with much higher resolution from WGS data. Surprisingly, this is also the case for a number of signatures generally attributed to hyper-mutation which now can be also detected at lower mutational burden in whole-genome sequenced samples. For example, whole-genome sequencing allowed identifying low exposures of SBS4, attributed to tobacco smoking, in the lung cancer genomes of light or former smokers [12].

Figure 2. The prevalence of mutational signatures of single base substitutions across samples profiled using whole-genome or whole-exome sequencing approaches.

A) Comparison of reported signature counts of WGS and WES samples from PCAWG and TCGA datasets. A positive log odds-ratio > 1 reflects signatures enriched in WGS, a log odds-ratio between 1 and −1 reflects signatures equally enriched in both sequencing technologies, and a log odds ratio < −1 reflects signatures enriched in WES. Any signature association with a −log10(Q-value) ≥ 2 is considered statistically significant.

B) Mutational profiles of signatures enriched in WES samples with annotated etiologies.

C) Mutational profiles of signatures enriched in WGS samples with annotated etiologies.

In addition to the systematic limitations in the mutational signatures identified from WGS and WES data, differences in the signatures identified between the two technologies can stem from the differences in the topographies of mutational signatures where mutational signatures depleted at nucleosomes or enriched at activator histone marks imply that these signatures have mutations strikingly at promoters and enhancers which will not be detected from WES. Despite the limitations of signature extraction, the low cost of WES makes it a preferred sequencing approach for signature extraction in research settings. To extend the utility of mutational signatures in the clinical setting, mutational signatures should be identifiable in sequencing data from targeted gene panels. Unfortunately, detecting mutational signatures from next-generation sequencing of targeted panels, routinely used in clinical practice, has proven to be a significant challenge.

Next-generation Sequencing of Targeted Panels of Known Cancer Genes

Targeted panel sequencing is a widely adopted NGS technique in clinical settings for quick, efficient, and affordable deep sequencing of a customized panel of cancer genes. Several gene panels, such as Memorial Sloan Kettering Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT) [58] and FoundationOne CDx [59] conduct targeted sequencing of exons and selected introns for 410 and 324 genes, respectively. These targeted gene panels have been instrumental in identifying an array of mutations, including single base substitutions, indels, copy number changes, structural variations, and fusions in cancer genes. More importantly, they allow for a reliable quantification of important biomarkers such as microsatellite instability (MSI) and tumor mutation burden (TMB), which are predictive of a patient’s response to immune checkpoint inhibitors and PD-1/PD-L blockade immunotherapy [60].

To extend the utility of mutational signatures as potential biomarkers in clinical settings, efforts have been made to extract mutational signatures from panel sequenced cancers, however, with limited success. Zehir et al. (2017) were only able to assign mutational signatures to ~9% of the total 10,000 MSK-IMPACT panel samples that had a very high overall mutation burden (≥ 13.8 mutations/Mb) [61]. In general, it is challenging to accurately decompose a small number of panel mutations into mutational signatures; the decomposition step using nonnegative matrix factorization and the signature assignment step using nonnegative least squares are both highly sensitive to low mutation counts and are likely to misassign mutational signatures [13]. Thus, the misattribution of mutational signatures can introduce many false positives. In general, it is important to note that panel sequencing has very similar limitations to those of Sanger capillary sequencing, which makes it infeasible for directly extracting mutational signatures [46].

In recent years, a shift towards developing supervised machine-learning methods for detecting specific mutational signatures from targeted sequencing technologies has gained traction. Gulhan et al. (2019) proposed Signature Multivariate Analysis (SigMA), a technique that allows to infer a specific clinically relevant mutational signature from targeted panel sequencing [23]. SigMA was specifically developed to detect signature SBS3 with the ultimate goal to identify homologous deficient patients who will benefit from treatment with PARP inhibitors. SigMA infers the presence and the prevalence of SBS3 in panel data by replacing the decomposition step with a clustering approach in the mutational signature extraction process. It uses a likelihood-based similarity metric to assign a sample, based on its mutational profile, a cluster of mutational signatures from a predefined set of tumor type-specific mutational signature clusters [23].

It is important to note that SigMA is only applicable to gene panels with at least five mutations. Therefore, it does not apply to cancer types with a low mutational burden, such as Ewing sarcomas or medulloblastomas [23]. Additionally, given SBS3 is a flat and featureless signature, there is a high likelihood of misassigning mutations to SBS3 from other more prevalent processes with correlated features such as clock-like signatures, that can greatly impact the sensitivity of SigMA in predicting SBS3 status. Despite these limitations, SigMA was applied to 383 cell lines from Cancer Cell Line Encyclopedia (CCLE) to predict SBS3 status. Cell lines classified as SBS3⁺ by SigMA had higher sensitivity (low IC50) to PARP inhibitors than cell lines classified as SBS3⁻ [23]. The results imply that SigMA performed very well in accurately predicting SBS3⁺ status for cell lines with good response to PARP inhibitors and SBS3⁻ status for cell lines with poor response to PARP inhibitors. Currently, SigMA is being employed to detect mutational signatures from other gene panel data for determining response to PARP and PD-L1 across ovarian, lung, and melanoma cancer patients [62,63]. In the future, carefully designed prospective trials with clinical endpoints such as RAD51 immunohistochemistry score (functional readout of HR proficiency) and Ki67 expression (prognostic marker for proliferation) after PARPi treatment will be required as ground truths to validate the clinical utility of SigMA and benchmark its performance with other HRD prediction approaches. The development of signature extraction methods based on supervised machine-learning approaches, such as SigMA, provides an exciting opportunity to extend the utility of mutational signatures in clinical settings. However, many of these methods are in their infancy and need further development, testing and validation before being formally deployed for routine use in the clinic. Such methods need to be robust against sequencing noise and the inherent variability in mutation counts across different cancer types, especially when working with data from panel-sequenced samples.

In recent years, there is a growing interest to replace targetted panel sequencing of a few cancer genes with Reduced Representation Sequencing methods (RR-Seq) for extracting mutational signature. RR-Seq allows sequencing a subset of the genome in a random and unbiased manner [64]. RR-Seq surveys larger parts of the genome, thus, capturing greater number of somatic mutations and providing more reliable data for extracting mutational signatures.

Reduced Representation Sequencing

Restriction-enzyme based sequencing (RAD-seq) is a reduced representation sequencing technique commonly employed in molecular ecology to study genomic evolution [64]. RAD-seq makes use of restriction enzymes to cut enzyme-specific sites in the genome. Therefore, the amount of the genome sequenced depends on the type and the number of restriction enzymes [64]. RAD-seq is not limited to a specific genomic region since the cut sites are present in both the intronic and exonic regions [64]. Over the past decade, the technique has been streamlined to use two enzymes for double digestion to reduce PCR duplicate errors and increase the multiplexing capabilities. In-silico methods such as ddradseqtools [65] are also used to predetermine the best combination of restriction enzymes and fragment lengths before conducting library preparation and sequencing [64–66].

In recent years, RAD-seq based techniques have been utilized to study cancer evolution and progression across tumor lineages [64]. Perner et al. (2020) took the technique a step further to identify mutational signatures using mutREAD (mutational signature detection by restriction enzyme-associated DNA sequencing) [67]. mutREAD is proposed to be a more accurate, cheaper, and time-efficient alternative to other NGS technologies. It allows to sequence a subset of the genome in an unbiased manner to extract mutational signatures from low-quality formalin-fixed paraffin-embedded (FFPE) samples. Perner et al. (2020) showed high concordance between signatures derived from PCAWG whole-genome sequencing data and signatures derived using mutREAD. They reported higher cosine similarity between mutREAD and WGS signatures in comparison to the cosine similarity between simulated WES and WGS signatures [67]. The higher concordance between WGS and mutREAD rather than WGS and WES-extracted signatures is likely due to similar mutational profiles of WGS and mutREAD signatures that contain mutations in both the coding and non-coding regions of the genome. On the contrary, WES signatures are biased for coding mutations, therefore, their mutational profiles do not closely match that of WGS derived signature. The high concordance between mutREAD derived mutational signatures and WGS signatures makes RAD-seq a promising sequencing technology for future extraction of mutational signature. Contingent on extensive validation, RAD-seq can facilitate incorporating mutational signatures as biomarkers in clinical settings given its low expected cost and multiplexing capabilities.

RNA and Single-cell RNA Sequencing

RNA sequencing (RNA-Seq) technologies are commonly used to determine the gene expression profiles, gene-splicing patterns, and cell states/types prevalent in cancer genomes. Only a handful of studies have used RNA-Seq for detecting somatic mutations due to the dependency on the gene expression levels for variant detection. Mutations in under-expressed genes with few transcripts will have low mutant allele frequency (MAF), making it hard to detect them with high accuracy. However, RNA-Seq analysis can help detect mutations in low purity samples since the MAF for variants in over-expressed genes is generally higher in RNA-Seq than DNA-based sequencing technologies like WES. Wilkerson et al. (2014) developed UNCeqR, a tool that integrates WES and RNA-Seq to detect variants in cancer genomes. UNCeqR captures a large fraction of mutations in low purity samples compared to mutations identified by variant detection techniques based only on either DNA or RNA-Seq. Such methods are crucial for variant discovery and subsequent mutational signatures in low purity samples [68].

RNA-seq is also instrumental for mapping and characterizing the somatic mutational landscapes of normal and precancerous samples. Over the past couple of two decades, the greater need for initiatives like Pre-cancer Genome Consortium (PCGC) has driven a paradigm shift in research with a greater focus on cancer prevention [69]. To develop better preventative methods and tackle cancer early on, interest in characterizing the genomic landscape of normal and precancerous lesions is gaining traction. The field has shifted from analyzing and identifying mutations in known cancer genes in a few normal macroscopic somatic clones in skin, esophageal, and breast samples [70–72] to making use of 6,700 normal samples [73] to comprehensively characterize the somatic landscape of 29 normal tissues [74]. In part, such efforts have been made possible by the development of techniques to detect mutations in RNA. Yizhak et al. (2019) developed RNA-MuTect to identify somatic mutations in normal RNA-Seq samples, making it possible to extract and characterize the mutational landscape across thousands of normal samples [74]. However, this study fell short of delivering the complete repertoire of the mutational signatures found across all 29 normal tissues. The low mutation counts for most samples only made it possible to characterize the mutational signature landscape for cancer types with high mutational burden. In the future, improved computational methods are needed for better inferring mutational signatures across different types of RNA-Seq normal tissues while accounting for differential rates of mutagenesis and controlling for age, microenvironment, tissue architecture, and tissue proliferation rates.

Single-cell RNA sequencing (scRNA-Seq) can help address the current limitations of RNA-Seq technologies in reliably deciphering the mutational landscape across and within different tissue types. Since the advent of single-cell sequencing technologies, scRNA-Seq has been instrumental in identifying transcriptionally distinct cellular populations [75], trace cancer cell lineages [76], and unravel the sub-clonal landscapes of tumors [77]. In the context of mutational signatures, scRNA-seq provides a promising avenue to identify expressed mutations in RNA transcripts to uncover ongoing mutational processes operative across distinct cellular populations in a tissue. Specialized tools have been developed that robustly identify single-cell copy number alterations [78] which provides an exciting opportunity to detect copy number signatures in scRNA-Seq data. However, technical artifacts introduced during the sequencing protocol greatly impact our ability to reliably detect single base substitutions at a single-cell resolution [79]. The Whole Transcriptome Amplification (WTA) step, which is required to amplify the small quantity of RNA for sequencing, can lead to variable genome coverage and introduce an array of technical artifacts that distort the mRNA expression levels [79]. The lack of specialized somatic mutation calling tools for scRNA-Seq that account for technical artifacts limits our ability to reliably extract mutational signatures from RNA at a single-cell resolution.

Single-cell DNA sequencing technologies grapple with similar challenges faced by scRNA-Seq. Nevertheless, a number of steps have been taken to remove technical noise from downstream mutation analyses of single-cell DNA sequencing for both variant detection and analysis of mutational signature.

Single-cell DNA Sequencing (scDNA-Seq)

To validate the proposed etiology of mutational signatures, one must perform proof of principle biological experiments to establish the causal link between the proposed molecular/biological processes and the acquisition of mutations attributed to specific mutational signatures. Several groups have experimentally validated mutational signatures across a diverse set of model systems, including human pluripotent stem cells, isogenic models, and C. elegans [17,80,81]. Mutational signatures are experimentally derived by repeatedly exposing parental clones to genotoxins or by knocking out genes of interest using CRISPR KO screens. The next couple of steps involve culturing the parental clones over several cellular generations to accumulate sufficient mutations, followed by growing multiple single-cell subclones over extended periods to generate ample cells for whole-genome sequencing [82]. Traditionally, germline variants in samples without a matching normal control were removed using panel of normals (PoN) and population genomic databases prior to any signature extraction [22,83]. However, a significant fraction of residual low-frequency/private germline variants has remained [22]. Levatić et al. (2021) propose an ancestry-matching approach to remove residual germline variants by assigning cancer cell lines a matching ancestry and subtracting its 96 mutational profile from the median 96 mutational profile of its ancestry-matched normal samples to yield the final somatic mutational profile for reliable signature extraction [84].

To overcome the laborious task of cell culture and subcloning, single-cell DNA sequencing can be employed. Single-cell DNA sequencing completely replaces the need for subcloning and can interrogate multiple cells from the same experiment, thereby providing sufficient statistical power to accurately discern mutational signatures from experimental models [82]. Despite the obvious advantage of using single-cell WGS in experimental settings, only a handful of studies have employed it to extract mutational signatures [22,85]. The field of single-cell genomics is still relatively new and is grappling with many technical issues that can confound the variant discovery and signature extraction process. Sequencing biases and artifactual errors such as C:G>T:A transitions can arise during the DNA denaturation and cell lysis step of the sequencing protocol [86]. Additionally, the whole-genome amplification (WGA) step, which is required to amplify single-cell genomes before sequencing, can introduce an array of artifactual errors depending on the type of WGA method used [87,88]. Multiple displacement amplification (MDA), which is the most commonly used WGA method, uses random primers and high fidelity polymerases to amplify the genome of a single cell [87]. MDA ensures high genome coverage but leaves behind random stretches of non-amplified regions in the genome leading to highly variable genome coverage, allelic imbalances, and drop-offs [87,89]. Thus, it is likely that real mutations in the single-cell data may be missed by variant callers used in bulk sequencing that have stringent variant allele frequency cut-offs. To account for amplification and artifactual errors, several single-cell variant callers such as SCAN-SNVs [89], SCaller [88], and MonoVar [90] have been developed. These tools have higher sensitivity and low false-discovery rates when calling mutations from single-cell data in comparison to mutation callers used for bulk sequencing [88–90]. Thus, when calling mutations from single-cell data, one needs to select the appropriate variant callers that account for sequencing and genome amplification artifacts.

Nonetheless, the mutational signature extraction process does not require all mutations of a cancer genome to reliably identify signatures [67]. Moreover, the signature extraction step itself can identify patterns of mutations attributable to WGA in single-cell analysis. Petljak et al. (2019) identified two WGA artifactual signatures, across 32 single cells from 16 different cancer cell lines, which were particularly enriched for C:G>T:A mutations in the GpCpN context. The mutations attributed to these signatures were mostly removed after filtering for mutations with a variant allele frequency of less than 50% [22]. Similarly, Lodato et al. (2018) identified three distinct single-cell mutational signatures associated with aging, the developmental stage, and the disease stage of the post-mitotic human neurons after extensive filtering of C:G>T:A mutations [91]. Some of the signatures reported in Lodato et al. (2018) were subsequently classified as MDA artefactual signatures by the same authors [92].

While promising strides are being made to remove single-cell sequencing artifacts, such artifacts can confound analysis of mutational signatures by masking the true mutations in a single cell. Therefore, there is a need to adopt alternative sequencing methods, such as single-molecule barcode sequencing, that can accurately identify both clonal and sub-clonal mutations with high sensitivity and specificity compared to single-cell sequencing.

Single-molecule Sequencing with Molecular Barcoding (SM-Seq)

Duplex sequencing or single-molecule sequencing is an NGS technology that sequences barcoded DNA molecules. The technology works by assigning a unique molecular barcode or unique molecule identifier (UID) to a DNA template molecule [93]. Once the tagged DNA molecules are amplified using PCR, DNA molecules with the same UID can be grouped into UID families. Mutations that exist in the majority of molecules in the same UID family are considered true mutations. This approach assumes that mutations, which existed on the original template DNA molecule before amplification, will be present on most copies of the original molecule. In contrast, artifacts that arose during amplification or sequencing will be found only in a small subset of the molecules in a UID family [93]. To summarize, duplex sequencing reliably identifies rare mutations and filters out the majority of artifactual mutations that are generated during library preparation and sequencing.

To maintain the low error rates of sequencing, duplex sequencing requires redundant sequencing to produce sufficient copies of DNA molecules for consensus variant detection. This step can be costly, especially when analyzing an entire cancer genome. Bottleneck sequencing system (BotSeqS) was developed to cut down on the need for redundant sequencing. BotSeqS performs a dilution step that retains a few double-stranded DNA molecules for subsequent PCR amplification and sequencing. BotSeqS is a successful technique applied to identify mutational patterns in normal samples exposed to environmental carcinogens and samples harboring defects in DNA repair pathways [94].

Despite the high accuracy of BotSeqS, it still grapples with end-repair artifacts. Methods that rely on mechanical fragmentation to shear genomic DNA for library preparation produce uneven blunts at the ends of a DNA template. DNA polymerases that repair these blunts occasionally introduce mutations in the terminal ends at specific NpCpA and NpCpT trinucleotide contexts [95]. PECC-seq (Paired-end and complementary consensus sequencing) is a method that uses a PCR-free consensus protocol to limit PCR artifacts and to computationally remove end-repair artifacts from samples. Even after bioinformatically filtering end-repair mutations from the terminal ends of the DNA template, it is still possible that such mutations can also exist in non-terminal regions of the DNA molecule. Therefore, a better experimental design is required to either perform mild mechanical fragmentation of the whole-genome molecule or to use high-fidelity enzymes to lower the error rates of single-molecule sequencing methods [95].

Nanorate sequencing (NanoSeq), a recently published duplex sequencing protocol, is a highly accurate method with an error rate of less than five errors per billion base pairs [96]. It can achieve a two-fold lower error rate than the rate of somatic mutagenesis in healthy cells by experimentally and computationally removing artifacts prevalent in BotSeqS. NanoSeq reduces end-repair artifacts by using DNA restriction enzyme-based fragmentation and the method uses non-A dideoxynucleotides (ddBTPs) to reduce errors from nick extension and a streamlined bioinformatics pipeline to remove artifacts introduced during the library preparation. Given its low error rate, NanoSeq will be a feasible sequencing technique for establishing the somatic mutational signature landscape of dividing and non-dividing/post-mitotic cells in the human body. It can be used for conducting large-scale in vitro studies to identify the mutational footprint of different mutagens across different tissues over time [96]. In the future, large-scale studies will need to be conducted to assess the applicability of NanoSeq for signature extraction. The use of a restriction-enzyme based DNA fragmentation provides a limited genome coverage which may be sufficient for reliable signature extraction if the restriction cut sites of the enzyme are randomly distributed in sufficient regions of the genome.

It is important to note that duplex sequencing methods do not sequence a single cell; instead, the method generally interrogates multiple molecules coming from many different cells. Thus, the mutational profile of duplex sequencing is usually an average somatic mutational profile across a population of cells. In recent years, novel methods such as acoustic cell tagmentation are developed that perform single-cell sequencing at a single molecule resolution [97]. Such methods will be instrumental in reliably extracting mutational patterns and mutational signatures at a single-cell resolution.

Long-read Sequencing

Long-read sequencing technologies involve sequencing continuous stretches of DNA ranging from 10kb to several Mbs. These technologies have proven instrumental for unravelling previously undetected complex structural variations (SVs) and providing the first telomere-to-telomere chromosome assembly [98]. Long-read sequencing technologies have also aided in detecting genomic variation and mutagenesis in complex regions, including telomeres, GC-rich regions, and acrocentric genomic regions with an abundance of tandem repeat sequences [98]. This is particularly exciting for researchers who aim to decipher the complete landscape of structural variation signatures, given the majority of SV signatures −16 in the recent PCAWG analysis and 6 in the 560 breast genomes – have been identified using exclusively short-read whole-genome sequencing [15,16]. The 300 bp single-read or the 150 bp paired-end read length sequenced by current NGS technologies, such as ones offered by Illumina, is too short for accurate detection and assembly of many SVs. Therefore, it is likely that short-read sequencing approaches currently miss the detection of more than 70% of the SVs that span more than 50 bps in the human genome [98]. With the availability of synthetic long read sequencing approaches such as Hi-C [99] and linked-read sequencing [100], which use short-read sequencing to synthetically generate long reads, our ability to detect SVs has dramatically improved. However, given the underlying technology for these synthetic methods is short-read sequencing, we are still limited in our ability to accurately detect all SVs. Single-molecule real time sequencing [98] can allow overcoming most of the issues in detecting SVs in cancer. Highly accurate long-read sequencing (HiFi) can now deliver long reads with accuracy rates comparable to the ones of short-read NGS technologies. Although not formally adopted in research and clinical settings, HiFi reads will ultimately revolutionize the detection of SVs in complex regions and pave the way for the identification of new signatures of structural variation.

CONCLUSION

In this review, we examined the relevance, limitations, and prospects of using different NGS technologies for exploring mutational patterns and for deciphering mutational signatures (Table 1). The biological question of interest, tumor type, cellular resolution (single-cell sequencing versus bulk sequencing), and the profile of mutational signatures should determine the use of appropriate NGS technology for extracting mutational signature. For each sequencing technology, it is important to be aware of the inherent sequencing biases and technical artifacts in order to apply appropriate filters before conducting mutational signature extraction. Similarly, it is also essential to be aware of the expected tumor mutational burden (TMB) of a sample before sequencing. TMB directly impacts the confidence in reliably extracting mutational signatures and can be a point of concern for low TMB samples when deciding to use NGS technologies that provide a reduced representation of the somatic mutation catalog of a cancer sample. Despite known limitations, the low cost of WES and targeted panel sequencing makes them a popular sequencing method routinely used in research and clinical settings. Additionally, the development of novel supervised machine-learning signature extraction methods for panel sequencing provide an exciting opportunity to extend the utility of mutational signatures into clinical settings and to use mutational signatures as potential biomarkers. However, many of these methods are in their infancy and need further development, testing and validation before being formally deployed in a clinical context. It is important to note that NGS data also provide the ability to acquire digital readouts of other molecular features in cancer genomes. Incorporating information about the regulatory and epigenetic landscape of cancer genomes can significantly extend our current understanding of signatures of unknown etiologies. With the decreasing costs for whole-genome sequencing and single-cell technologies, it will become possible to integrate multiple single-cell data modalities, including mutational signatures, chromatin accessibility, and RNA expression to provide valuable insights into the causal mechanisms driving known and unknown mutational signatures.

Table 1:

Significance and limitations of the use of different sequencing technologies for detecting mutational signatures.

Sequencing Technology	Sequenced Region(s)	Significance	Limitations
Next-generation Sequencing of Targeted Panels of Known Cancer Genes	Exons and introns of selected gene(s)	Capillary Sanger sequencing was instrumental in identifying the first set of mutational signatures in TP53 across different tissue types. Driving the development of machine learning based signature extraction methods such as SigMA that replace the decomposition step with a clustering step to infer mutational signatures in panel sequenced data.	Allows identification of single base substitution signatures in highly mutated samples. Difficult to assign mutational signatures to a single sample due to low mutation burden. Mutations may be under strong selection pressure in coding regions.
Restriction-enzyme based sequencing (RAD-Seq)	Regions with restriction enzyme cut sites	Streamlines panel-based sequencing approach to include a more unbiased representation of mutations from both the coding and non-coding regions of the genome using an optimal combination of restriction enzymes.	The approach has not been widely adopted in research or clinical settings. Limited to detecting mutational signatures in bulk tissues. Not able to reliably detect other molecular features of cancer genomes (e.g., mutations in driver genes).
RNA Sequencing (RNA-Seq)	Transcribed regions	RNA-Seq has been used for mutational signature analysis in normal samples. Mutation calling on RNA-Seq profiles of more than 6,700 normal samples provided the comprehensive somatic mutation landscape for 29 normal tissue types and the prominent mutational signatures in the skin and lung normal tissues.	Difficult to identify mutations in under-expressed transcripts and/or genes. Biased mutational profile specific to transcribed regions of the genome limits resolution and mutational burden needed for reliable signature extraction.
Whole-exome Sequencing (WES)	All exons	WES was instrumental in extracting and establishing the repertoire of single base substitution (SBS) signatures across 7,000 cancers across 40+ tumor types. Cheap and reliable sequencing technology for identifying signatures in hyper-mutated cancers such as skin-melanoma, lung cancers, etc.	Variable sequencing depth due to GC-rich regions and sequencing artifacts can lead to wrong signature assignment and attributions. Low mutation burden limits the identification of signatures with ‘flat profiles’ such as SBS3 and signatures for large-scale events such as indels, doublet-base substitutions, and structural variations.
Whole-genome Sequencing (WGS)	Genome	WGS provides a complete somatic mutational catalog that enables the extraction of both small-scale and large-scale mutational signatures such as signatures of single-base substitutions, indels, doublet-base, and structural variations. Reliable sequencing technology for identifying signatures in low tumor mutation burden (TMB) cancers and signatures with ‘flat mutational profiles’ such as SBS3, SBS8, SBS5, etc.	WGS is more expensive that other sequencing approaches. WGS is not routinely used in clinical settings.
Single-cell Sequencing (SC-Seq)	Genome of a single cell	SC-Seq provides a better resolution of mutational processes operative at a single-cell level. Aids in unwinding the heterogeneity of mutational processes prevalent at the bulk tissue level.	Limited suitability for extracting/validating mutational signatures in normal tissues or in experimental models due to large number of whole-genome and -transcriptome amplification artifacts and variable sequencing depth.
Duplex Sequencing	Parts of the genome	Duplex sequencing accurately identifies both clonal and none-clonal mutations with low variant allele frequency. Appropriate sequencing technology for studying mutational signatures in different experimental models and examining mutational signatures in normal tissues.	End-repair sequencing artifacts produced by the mechanical fragmentation step of the library preparation can lead to spurious signature assignments. Not routinely used in clinical or research settings.
Long-read Sequencing	Genome	Long-read sequencing increases the accuracy in detecting structural variations not captured by short-read sequencing methods.	High sequencing error rate and/or high sequencing cost. Lack of established consensus variant calling methods for long read sequencing.