Detection of structural DNA variation from next generation sequencing data: a review of informatic approaches

Abstract

Next generation sequencing (NGS), or massively paralleled sequencing, refers to a collective group of methods in which numerous sequencing reactions take place simultaneously, resulting in enormous amounts of sequencing data for a small fraction of the cost of Sanger sequencing. Typically short (50–250 bp), NGS reads are first mapped to a reference genome, and then variants are called from the mapped data. While most NGS applications focus on the detection of single nucleotide variants (SNVs) or small insertions/deletions (indels), structural variation, including translocations, larger indels, and copy number variation (CNV), can be identified from the same data. Structural variation detection can be performed from whole genome NGS data or “targeted” data including exomes or gene panels. However, while targeted sequencing greatly increases sequencing coverage or depth of particular genes, it may introduce biases in the data that require specialized informatic analyses. In the past several years, there have been considerable advances in methods used to detect structural variation, and a full range of variants from SNVs to balanced translocations to CNV can now be detected with reasonable sensitivity from either whole genome or targeted NGS data. Such methods are being rapidly applied to clinical testing where they can supplement or in some cases replace conventional fluorescence in situ hybridization or array-based testing. Here we review some of the informatics approaches used to detect structural variation from NGS data.

Introduction

The detection of structural DNA variation has long played a role in the diagnosis of cancer and Mendelian disorders, predating the advent of modern DNA sequencing (1,2). Structural DNA variation is generally defined as variation in a DNA region larger than 1 kb and includes several classes such as translocations, inversions, insertions/deletions (indels) and copy number variations (CNVs) (3). In the clinical laboratory, the detection of structural variation is performed by a diverse group of methods. Among the oldest and most basic methods for structural variation detection is routine cytogenetics, in which metaphase chromosomes are stained and morphologically evaluated by light microscopy. Conventional cytogenetics represents an unbiased approach for the detection of translocations, inversions, and large deletions or insertions; however, most clinical cytogenetic as-says are performed at the 350–500 band level and are of limited resolution and sensitivity. For example, clinically relevant events such as the FIP1L1-PDGFRA deletion on chromosome 4q12 in myeloid neoplasms, unusual or multi-partner rearrangements, and variants present in less than 5% of cells are generally not identified by conventional cytogenetics (4–6). Another major limitation of conventional cytogenetics is the requirement for cultured metaphase cells, which are generally not obtainable in solid tumors. Fluorescence in situ hybridization (FISH) offers considerable advantages over conventional cytogenetics, including increased resolution, the ability to test fixed interphase cells, faster turnaround time, and greater sensitivity. For solid tumors, FISH is often the method of choice for the detection of recurrent mutations, such as ALK rearrangements in lung cancer, MYCN amplification in neuroblastoma, and 1p/19q deletions in oligodendrogliomas (7–9). While FISH offers improved sensitivity compared with that of conventional cytogenetics, the evaluation of multiple loci requires multiple probes and FISH assays to be run, increasing the complexity of testing. DNA microarray technology has proved to be another reliable clinical method for the detection of structural variation, especially CNV and loss of heterozygosity (LOH). However, unlike FISH, DNA microarrays are unable to detect balanced translocations (10,11).

Next generation sequencing (NGS), often referred to as massively paralleled sequencing, is a collective group of methods characterized by their high sequencing throughput (12). Currently available NGS platforms include the Illumina HiSeq/MiSeq, Life Technologies Ion Torrent/Ion Proton, Life Technologies SOLiD, and Roche 454. In contrast to Sanger sequencing, which produces a single long (often >1 kb) read using dye terminator chemistry, NGS methods typically generate millions of short reads on the order of 50–250 bp using reversible sequencing chemistries (13). NGS methods have allowed for unprecedented discovery in cancer, including acute myeloid leukemia, lung cancer, and breast cancer, and are now being applied in the clinical setting for evaluation of cancer predisposition syndromes, developmental delay, and cancer prognosis (14–19). NGS may be used to generate whole genome data, generate exome data (all coding sequences in the genome), or target specific genes or loci of interest (20). While whole genome data is generally low coverage (8–30 × coverage) and suitable for the detection of constitutional variants, by targeting sequencing to specific genes or regions of interest, coverage may be increased to 1,000× or higher, permitting more sensitive evaluation of gene variants and subclonal populations in cancer (21).

NGS-based diagnostics are rapidly becoming part of the clinical genomic testing and are now routinely offered by many commercial and academic laboratories. One of the key features of NGS-based diagnostics is its ability to detect a full range of genetic variation, offering the potential to greatly streamline testing by using a single analysis platform. For example, prognostic evaluation of acute myeloid leukemia generally requires the use of multiple technologies including PCR and fragment sizing to detect FLT3 internal tandem duplications and NPM1 insertions, Sanger sequencing to detect CEBPA, IDH1/2, and DNMT3A mutations, and FISH to detect MLL, RARA, CBFB, and RUNX1 rearrangements. Such complex evaluations require numerous highly trained personnel and are often prohibitively expensive. NGS-based testing, however, can identify SNVs, insertions, and trans-locations in a single assay, often for considerably lower cost compared with that of conventional workups (22,23).

Here we review methods for the identification of DNA structural variation by NGS, with particular emphasis on methods suited for targeted sequencing likely to be used in the clinical laboratory. This review will focus on three major types of structural variation: translocations, CNV, and insertions/deletions. While numerous software tools are available for NGS analysis, currently no single tool is capable of identifying the full range of DNA variation, and we will review some of the most widely used, publicly available software packages.

NGS informatics

Algorithms for detection of structural variation from NGS data rely on one or more of the following: discordant paired-end reads, split reads, or depth of coverage. Discordant paired reads are read pairs that do not map together in the ordinary way: The paired ends may map to different chromosomes or to the same chromosome either in the incorrect orientation or in the proper orientation but, for instance, too far apart in the chromosome. Split reads are single reads that map to the genome discontinuously: The first part of the read maps to one genomic region and the remainder to another. Because of the short read lengths currently available from NGS data, split reads are most useful and reliable from paired-end data, in which one end maps uniquely to the genome, serving as an “anchor,” and the other end is a split read. Finally, the depth of sequencing coverage local to a particular point in the genome provides evidence of structural variation. While changes in read depth over large regions often indicate copy number changes, more subtle variation in sequence coverage is often seen near the breakpoints of other types of structural variation.

The performance of any method for detection of structural variation depends critically on the type of sequencing data available. For instance, split-read methods to detect trans-locations generally require adequate coverage so that the translocation breakpoints are spanned by several split reads, and they will not perform well using low coverage whole genome sequencing data. Similarly, indels can be detected from exome (or targeted-capture) data using paired- or split-read methods only if at least one of the breakpoints falls within or near the captured regions. Finally, any method that relies on read depth will perform differently for whole genome as compared with exome or targeted-capture data, as the depth of coverage in targeted-capture data is particularly susceptible to GC bias, uneven coverage near the boundaries of the capture baits, and other systematic biases. A summary of tools used for structural variation detection is presented in Table 1.

Table 1.

Software tools for evaluation of structural variation in NGS data

	Comment	Download link
Translocations and Inversions
Discordant paired end
BreakDancer	Fast, simple to run	http://breakdancer.sourceforge.net
Hydra	Considers multiple mappings of discordant pairs	https://code.google.com/p/hydra-sv/
VariationHunter	Considers multiple mappings of discordant pairs	http://variationhunter.sourceforge.net/Home
PEMer	Simulates structural variants	http://sv.gersteinlab.org/pemer/introduction.html
GASVPro	Improved specificity by combining info from discordant pairs and coverage depth	http://compbio.cs.brown.edu/software.html
Split end read methods
CREST	Requires soft-clipped reads generated during alignment	http://www.stjuderesearch.org/site/lab/zhang
Slope	Replaced by ClusterFAST	https://github.com/eduncavage/clusterfast
CNV
Raw coverage based
EWT	Whole genome only, does not require normal control	http://rdxplorer.sourceforge.net
Coverage ratio based
SeqSeq	Whole genome only; requires normal control	http://www.broadinstitute.org/software/cprg/?q=node/39
CNVnator	Whole genome only; requires normal control	http://sv.gersteinlab.org
CNAseg	Whole genome only; requires normal control	http://www.compbio.group.cam.ac.uk/software/cnaseg
CNV-seq	Whole genome only; requires normal control	http://tiger.dbs.nus.edu.sg/cnv-seq/
CONTRA	Exome or targetd panels; requires normal controls	http://sourceforge.net/projects/contra-cnv/
CoNVEX	Exome; requires normal controls	ftp://ftp.sanger.ac.uk/pub/users/pv1/CoNVex/Docs/CoNVex.pdf
ExomeCNV	Exome or targeted panels; requires normal controls; evaluates B-allele frequency	https://secure.genome.ucla.edu/index.php/ExomeCNV_User_Guide
Insertions and Deletions
Pindel	Uses pattern-growth algorithm to find medium and large indels	http://gmt.genome.wustl.edu/pindel/current/
GATK	Full package for SNV and indel detection	http://www.broadinstitute.org/gatk/download
VarScan2	Detects SNVs and indels	http://varscan.sourceforge.net
Dindel	Applies probability-based filtering to reduce false positives	https://sites.google.com/site/keesalbers/soft/dindel
Stampy	Limits false-positive indel calls	https://wiki.gacrc.uga.edu/wiki/STAMPY
SAMtools	General purpose suite detects SNVs and indels	http://samtools.sourceforge.net

Detection of translocations and inversions

Many algorithms, including BreakDancer, Hydra, PEMer, and VariationHunter, for the detection of structural variation rely on the presence of discordant paired reads (24–27). In the case of interchromosomal translocations, one member of the pair maps to one chromosome and its mate to another (Figure 1). In the case of inversions or intrachromosomal translocations, the two ends map to the same chromosome but in the wrong orientation or the wrong distance apart. These algorithms are generally quite sensitive in detecting translocations and inversions in mappable areas of the genome; however, in general, they can detect breakpoints only with low resolution and often suffer from low specificity, particularly when one member of the pair maps to a repetitive region or to a region that shares homology with other areas of the genome. Furthermore, translocations, because of the mechanisms by which they are generated, tend to occur in regions with repetitive elements, such as tandem duplications and transposons (26). Thus, true positives exist in these regions and are difficult to discern from the many false positives. The Hydra and VariationHunter software packages attempt to detect structural variations occurring in such repetitive regions by considering multiple possible high scoring mappings per read, rather than just the unique, best mapping. Most paired-read methods for detection of structural variation rely on heuristic cutoffs to filter out false positives, such as the number of supporting read pairs. One recently described algorithm, however, GASVPro, combines paired-end and subtle coverage depth signals into a probabilistic model to achieve greatly improved specificity in detection of structural variation (28).

Figure 1.

Identification of translocations from discordant paired-end reads. (A) In this example, a t(4;11) translocation is identified by discordant paired-end reads. Read pairs are first identified, in which one end maps to the targeted region (in this case the MLL gene on 11q23) and the other end maps to a different chromosome. (B) Discordant paired-end ead methods are subject to high false-positive rates due to sequence-mapping errors and repeat regions in the genome. Most translocation identification software employs filtering criteria to reduce the number of false-positive calls.

In order to avoid the high false-positive rates inherent to most paired-read approaches, and to better localize the breakpoints, as is needed for orthogonal validation of the breakpoints by PCR, some algorithms for structural variation detection make use of split reads, in which a single read contains spans a breakpoint between two distant genomic regions. One indication of the presence of split reads in aligned sequence data is the existence of clusters of soft-clipped reads. Soft clips are produced by some alignment software (including Novoalign and BWA) when one end of a paired-end read maps uniquely and entirely to the genome but the other end does not. If the second end maps only partially, but in the correct orientation and has an [insert size] within the normal range, the remainder of the sequence is represented as a “soft clip” (29). These soft-clipped reads often indicate reads with split mappings and so can be used to localize translocations with single-base accuracy. CREST and ClipCrop are two algorithms that make use of soft-clipping information to detect split reads (30,31). SLOPE is another method for detection of translocations by split reads (32). Instead of using soft clips, however, it performs a local realignment of the unmapped “orphaned” mates in the vicinity of the uniquely mapped mate and looks for single breakpoints supported by many split reads.

Detection of copy number variants

Copy number variants are defined as stretches of DNA, longer than a kilobase, that are present in the genome with an abnormal number of copies. These include large deletions and duplications, as well as unbalanced translocations.

In whole genome sequence data, large deletions are readily detected by the same methods as moderately sized deletions, that is, paired-end methods such as BreakDancer or split-read methods such as Pindel (25,33). Large deletions are easier to detect than smaller indels using paired-end methods, as they are easily distinguished from normal variation in the insert size. Large duplications are more difficult to detect, as there is no single read or read pair spanning the insertion. One software tool capable of detecting large insertions is Pindel, which uses a pattern-growth approach to detect breakpoints, at which the sequenced genome diverges from the reference. If two such breakpoints occur in the same chromosome with an appropriate orientation, Pindel can piece them together to discover a large insertion event.

Discordant paired-end and split-read methods will generally not perform well for CNV detection from exome or targeted-capture sequence data, as they require the capture of at least one of the breakpoints. In this case, algorithms that examine the sequence depth of coverage are the primary means for detecting CNV. Read depth methods for CNV can be applied to both whole genome and targeted-capture data but require different considerations (Figure 2). In the case of whole genome sequence data, the pattern of deletions and duplications is readily apparent from the sample's coverage profile, albeit at low resolution. The computational problems here are to accurately localize the breakpoints and to determine the number of copies present in each segment. Several software packages have been developed to address these issues; some of the existing algorithms are intended to detect CNV based on the read epth profile of a single sample, whereas others require a control sequence for comparison.

Figure 2.

CNV by DOC analysis. In this example, CNV is called by first obtaining the DOC for every position in the targeted sequencing region. Next, DOC data must be normalized, which can be accomplished by a number of approaches, including comparing to paired normal samples (in the case of cancer), pooled normal controls, or the mean sample coverage. Once coverage is normalized, regions of constant CNV are identified, and CNV calls are then made using a variety of probabilistic models.

Algorithms such as event-wise testing (EWT) for CNV detection from whole genome sequence data do not require control sequence data; rather, they rely on deviations in coverage depth from the sample's mean depth (34). Since many factors, including GC content, influence a sample's coverage profile, these methods must attempt to correct for these biases to provide adequate specificity (35). The general procedure for these algorithms is to divide the genome into nonoverlapping bins of equal size and then calculate the mean depth of coverage (DOC) for each bin. After the read depths are corrected for GC and other biases, a segmentation algorithm is used to divide the genome into regions of constant copy number.

Although methods for CNV detection with no control sample must explicitly account for GC bias, other methods, including SegSeq, CNVnator, CNAseg, and CNV-seq, designed for either tumor-normal or case-control comparison avoid this issue by comparing the same region (which should be subject to the same GC bias) across multiple samples (36–39). These approaches similarly partition the genome into regions, calculate the depth of coverage ratio between case and control for each region and then partition the region into segments of equal copy number, using a variety of approaches, including hidden Markov models (HMMs) and circular binary segmentation (40). These algorithms, because they rely on the coverage ratio rather than the raw coverage profile, permit finer mapping of CNV boundaries using, for instance, mean-shift approaches from signal processing (37).

Detection of CNV from exome or targeted-capture sequence data presents unique challenges due to the increased GC bias inherent to targeted-capture data, and the discontinuous nature of the coverage profile. For custom targeted-capture sequence data, in which large regions (>1 kb) are targeted, CNV can be detected within single samples (no controls) following correction for GC-content and edge effects (41). However, because of the small size of targets in typical exome-capture data, many current algorithms for CNV detection require either a paired normal sample or a panel of population controls. CONTRA, one such method for CNV detection from exome data, first calculates the tumor/ normal coverage ratio exome-wideband then employs a normal approximation to detect CNV at the exon level (42). Finally, exon-level deletions or duplications are merged into larger CNV using circular binary segmentation (CBS). CoNVEX detects CNV using a similar strategy, first denoising the coverage ratio using a discrete wavelet transform, and then identifying copy gains and losses via a hidden Markov model (43). A third, similar, approach is taken by ExomeCNV, which segments the exome into regions of equal copy number using CBS, based on the tumor/normal ratio (44). However, ExomeCNV also models the B-allele frequencies to detect LOH, which can be used to corroborate CNV calls, deletions in particular.

An additional category of CNV detection algorithms, designed to detect sporadic CNVs from population exome sequence data, uses principal components of the matrix of read counts, over samples and exons, to normalize the read count data (45,46). In the absence of recurrent CNVs, the top principal components, which explain the bulk of the exome-wide variance in DOC, should represent experimental noise, including batch effects and GC-bias. Thus, removing the top principal components (i.e., projecting the data onto the space defined by the remaining components) should eliminate these biases. It should be emphasized, however, that such methods are intended for detecting sporadic CNVs; recurrent CNVs tend to be picked up by the top principal components, and their signals are therefore lost in the process of normalization.

Detection of insertions and deletions

Indels are common in the human genome and contribute to genetic diversity and human disease (47–49). In the clinical molecular oncology laboratory, the detection of small (defined here as <10 bp) and medium (defined here as >10 but <1 kb) indels is important to many cancers. Of particular clinical significance are the NPM1 insertion, FLT3 internal tandem duplication (FLT3-ITD), KIT exon 8 indels in acute myeloid leukemia, and EGFR exons 19 and 21 insertions and deletions in lung cancer (50–53). While small-and mediumsized indels are usually simple to detect by Sanger sequencing or gel capillary–based sizing methods, indel detection by NGS methods has been challenging largely because of the short read lengths generated by NGS methods. In general, small indels can be called with reasonable sensitivity from NGS data, although the specificity tends to be low. Medium-sized indels, such as the FLT3-ITD, however, have proven difficult to detect by most, but not all, methods (22). Further, most indel detection software is biased to detect deletions over insertions, as inserted sequences are more difficult to align to the reference sequence as described below. Indel detection software can be divided into four major categories, although there is considerable overlap among software packages.

The most common indel detection methods are alignment-based methods that are generally optimized to detect small indels. These methods are often included in popular variant detection packages such as SAMtools, the genome analysis toolkit (GATK), or VarScan (54–56). Alignment-based methods generally rely on probabilistic models to make indel calls based on data obtained during the initial read mapping and alignment process (Figure 3A). For the example of a small insertion, reads containing the insertion are first mapped to the reference sequence using gapped alignments, a step generally performed by the read mapping software (BWA, Novoalign, etc.) (57). Indel variant detection software will then use the alignment data to call an indel event after applying a filtering step to differentiate common sequence alignment errors from true indels. There are numerous indel detection programs that rely on this method, including Dindel, Stampy, and others, in addition to the more general packages described previously (27,58–61). These methods differ principally in the model used to discriminate between alignment errors and true indel calls, often resulting in greatly discrepant indel calls between software and orthogonally generated data (62). Of the many alignment-based indel detection methods, no single program has proven to be completely accurate and all require considerable validation when used clinically. A major drawback of alignment-based methods is the requirement that indels be contained within a read and identified during the initial read mapping and alignment stage, limiting insertion detection to approximately 15% of the total read length (22,59). However, larger deletions may be detected by these methods; this is often referred to as “deletion bias.”

Figure 3.

Methods for indel detection. (A) In this example, a small insertion is identified by alignment-based calling methods. Such insertions are generally identified during initial read mapping and alignment and are evaluated by indel detection programs using different models to exclude false-positive results due to sequencing or read mapping errors. (B) A medium-sized insertion is identified by split read mapping methods. In this example, an insertion (red) present in the sequenced DNA is detected by first identifying paired-end reads in which one end maps and the other (containing the inserted sequence) does not. The inserted sequence is reconstructed from the overlapping, unmapped single-end reads. (C) An insertion detected by paired-end methods. In this example, the sequenced DNA contains an insertion and read pairs mapping to the flanking normal reference sequence show a shorter than expected distance between ends, allowing for an insertion to be inferred.

Split read mapping methods, such as the widely used Pindel program, are capable of identifying medium-sized indels that are often missed by alignment-based indel software (33). These methods, including de novo alignment and others, function by first identifying discordant paired-end reads in which one end maps completely to the reference sequence and the other end does not. The unmapped end of these ”single-end anchored reads” are then clustered or subjected to de novo alignment to determine the exact sequence of an insertion (Figure 3B). Using this approach, insertions longer than the read length can be identified, as such methods do not rely on the initial read mapping step. Split read mapping methods are particularly suited to identify clinically relevant indels, such as the FLT3-ITD. However, these methods are also subject to a higher false-positive rate, as they generally do not use probabilistic models to discriminate between alignment errors and true indel events.

Methods based on paired-end read mapping identify large indel events by comparing the expected distance between read pairs to the actual mapped distance. Such methods include PEMer, Hydra, and BreakDancer (25,27). For example, in the case of a 50-bp insertion, if the distance between read pairs is normally distributed, with mean 200 bp and then multiple pairs aligning to the same area with a distance between read pairs of approximately 150 bp would result in an insertion call (Figure 3C). Paired-end read mapping methods are therefore able to detect medium-sized insertions and deletions from mapped data. However, in most cases, the exact inserted or deleted sequence will not be known. Another major drawback of paired-end read mapping methods is that they are insensitive to small insertion or deletion events, owing to the difficulty in separating small perturbations in read-pair distance from the normal background variability. Another new class of indel-detection software is those based on machine learning methods in which insertions and deletions identified by various methods are filtered against empirically derived training set data to reduce the false-positive rate (63). These newer methods have yet to be rigorously tested but promise to reduce the inherent false-positive rate of indel detection, especially in homopolymer tracts and areas of low sequence complexity.

Conclusion

A full range of structural variation can be detected from NGS data, including translocations, CNVs, and indels. It is important to note, however, that there is currently no single informatic method capable of identifying the full range structural DNA variation, and multiple complementary tools are required for robust variant detection. Further, the use of any software for structural variation identification in the clinical laboratory will require extensive validation, as such methods perform differently depending on assay design (targeted vs. whole genome) and average DOC.