ChopSticks: High-resolution analysis of homozygous deletions by exploiting concordant read pairs

1471-2105-13-279 1471-2105 Methodology article ChopSticks: High-resolution analysis of homozygous deletions by exploiting concordant read pairs YasudaTomohirotyasuda@hgc.jp SuzukiShinshinout@hgc.jp NagasakiMasaonagasaki@megabank.tohoku.ac.jp MiyanoSatorumiyano@hgc.jp

Human Genome Center, Institute of Medical Science, University of Tokyo, 4-6-1 Shirokanedai, Minato-ku, Tokyo 108-8639, Japan

Department of Integrative Genomics, Tohoku Medical Megabank Organization, Tohoku University, 2-1 Seiryo-machi, Aoba-ku, Sendai 980-8573, Japan

BMC Bioinformatics

Sequence analysis (methods)

1471-2105 2012 13 1 279 http://www.biomedcentral.com/1471-2105/13/279 10.1186/1471-2105-13-27923110596 74201259201230102012 2012Yasuda et al.; licensee BioMed Central Ltd.This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Background

Structural variations (SVs) in genomes are commonly observed even in healthy individuals and play key roles in biological functions. To understand their functional impact or to infer molecular mechanisms of SVs, they have to be characterized with the maximum resolution. However, high-resolution analysis is a difficult task because it requires investigation of the complex structures involved in an enormous number of alignments of next-generation sequencing (NGS) reads and genome sequences that contain errors.

Results

We propose a new method called ChopSticks that improves the resolution of SV detection for homozygous deletions even when the depth of coverage is low. Conventional methods based on read pairs use only discordant pairs to localize the positions of deletions, where a discordant pair is a read pair whose alignment has an aberrant strand or distance. In contrast, our method exploits concordant reads as well. We theoretically proved that when the depth of coverage approaches zero or infinity, the expected resolution of our method is asymptotically equal to that of methods based only on discordant pairs under double coverage. To confirm the effectiveness of ChopSticks, we conducted computational experiments against both simulated NGS reads and real NGS sequences. The resolution of deletion calls by other methods was significantly improved, thus demonstrating the usefulness of ChopSticks.

Conclusions

ChopSticks can generate high-resolution deletion calls of homozygous deletions using information independent of other methods, and it is therefore useful to examine the functional impact of SVs or to infer SV generation mechanisms.

Background

Today, next-generation sequencing (NGS) technologies are essential tools in genome analysis, because they enable us to simultaneously obtain sequences of up to hundreds of billions of base pairs 1 . These technologies enable the characterization of not only small variations such as single-nucleotide polymorphisms (SNPs) but also large-scale mutations such as insertions, deletions, tandem duplications, and inversions. Mutations of these types are collectively called structural variations (SVs) and are frequently observed even in healthy individuals 2 3 4 . Because SVs affect a much larger portion of genomes than small variations, including SNPs, they have a great impact on biological functions.

Current NGS methods can sequence paired reads, which are pairs of reads several hundred bases away from each other. This ability is useful for analyzing SVs because paired reads can be aligned with the reference genome more accurately than single reads, and because we can analyze structures of genomes larger than the size of each read. However, SV detection is still a difficult task, because it requires analysis of the complex structures involved in an enormous number of alignments of paired reads with the reference genome, and because read sequences and alignments include unavoidable errors. Therefore, for example, a false detection rate (FDR) up to 10% had to be tolerated even when determining just the existence of each SV in the 1000 Genomes Project 2 . It is obviously more difficult to accurately detect the exact positions of SVs. Nevertheless, high-resolution SV calls are necessary to elucidate the functional impact of SVs and molecular mechanisms that generate SVs. Moreover, to conduct a large-scale analysis, SV detection methods for data with a low depth of coverage (hereafter simply referred to as coverage) are desirable, because whole genome sequencing is not easy even with NGS technologies.

Current methods for SV detection search for signatures that indicate SVs hidden in read sequences and their alignments with the genome sequences. The following are basic signatures used for SV detection 2 3 4 .

• Read pair (RP) 5 6 7 : If pairs of reads have aberrant strands or distances, they are likely to be caused by SVs. Such pairs are called discordant pairs, and normally mapped ones are called concordant pairs. If strands of a discordant pair are as expected, a larger distance than expected indicates a deletion, whereas a smaller distance indicates an insertion. There are several categories of methods that detect discordant pairs by using mapping distances.

• Threshold-based: A pair with a mapped distance larger or smaller than a predefined threshold is defined as a discordant pair. The threshold is μ±3 σor μ±4 σfor BreakDancer 5 and VariationHunter 6 where μand σare mean and standard deviation of mapping distances, or median fragment size ± 10 median absolute deviations for HYDRA 7 .

• Distribution-based: Although the mapped distance of a single pair might vary by tens or hundreds bases even without SVs, larger (smaller) mapping distances of many pairs in the same region indicate deletions (insertions). Such reads can be detected by statistical tests on the distribution of mapped distances 5 8 . Pairs detected in this way might have mapping distances more similar to the expected distance than those of other methods. Nonetheless, we still call them discordant pairs in this paper to unify the word used to refer pairs that support SVs.

• Graph-based: Recently Marshall et al. 9 proposed a new method CLEVER based on the graph theory. CLEVER constructs a graph where a node represents an alignment of a read pair and the genome, while an edge means that connected alignments potentially support the same allele. In this graph, a clique corresponds to a set of pairs supporting the same allele. CLEVER detects SVs by finding maximal cliques (max-cliques). CLEVER has an ability to find more than one max-clique overlaping each other, each of which supports a different allele. Therefore CLEVER can distinguish more than one SV located at the same locus, for example, two deletions of different sizes in a diploid genome.

• Read depth (RD) 10 11 : If coverage changes at some position in the genome, this indicates a copy number variation.

• Split read (SR) 12 : If an alignment of a read and the genome includes only a part of the read, this indicates a position of a breakpoint. Here, a breakpoint is the boundary between a region affected by some SV and its unaffected flanking region.

• Sequence assembly (AS) 7 13 : If the coverage is sufficient, assembling NGS reads around an SV reveals the exact sequence around the SV and the positions of breakpoints.

The most popular signature used to detect SVs is threshold-based RP. Methods based on this signature can detect SVs from a small number of discordant read pairs; therefore threshold-based RP methods can be applied to low-coverage data. However, threshold-based RP methods localize SVs only to regions surrounded by discordant read pairs, thus causing some ambiguity. For RD methods, the problem of resolution is much bigger. Because RD methods involve calculation of coverage in windows of a fixed size, its resolution cannot be finer than the window size. Methods based on the SR signature can determine positions of breakpoints up to base-pair-level (bp-level) resolution if there are reads covering the breakpoints. However, such reads might not exist, in particular when coverage is low, because of unevenness of coverage or repeat elements to which reads cannot be aligned uniquely. Moreover, because such a split alignment is shorter than a read itself, careful analysis is required to avoid spurious matches. If coverage is sufficiently high, AS methods would ultimately reveal the exact positions of SVs at bp-level resolution. Although extremely deep sequencing can be conducted by targeted sequencing 14 , it is still expensive to obtain paired reads of high coverage over the entire genome so that assembly can be performed. In fact, a previous study has indicated that the sensitivity of AS methods is rather low (Table S6B of Mills et al. 3 ).

Because these signatures have their own advantages and disadvantages, it is desirable to combine more than one method 4 . In fact, several methods that use more than one signature have been proposed recently 15 16 . In combined approaches, we should integrate SV signatures that are independent of each other. In this paper, we propose a new method called ChopSticks that improves the resolution of deletion calls for homozygous deletions generated mainly by threshold-based RP methods. ChopSticks is especially valuable when target SVs are expected to be homozygous as those of inbred mice whose genomes are homozygous at virtually all loci 17 . ChopSticks exploits positions of concordant read pairs in addition to those of discordant ones. Thus far, they have been ignored in threshold-based RP approaches, and therefore, our method can improve the resolution by using this new independent information. As explained below, ChopSticks is effective even for data whose coverage is low.

The organization of this paper is as follows. First, we theoretically analyze the improvement of the resolution achieved by exploiting concordant read pairs. Next, we present our computational method ChopSticks that improves the resolution of homozygous deletion calls. After that, we demonstrate the effectiveness of ChopSticks in computational experiments. Then, we present our conclusions. In addition, we illustrate details of our method and experiments in Methods section.

Results and discussion

Strategy for resolution improvement

Theoretical estimation of resolution

Here we present results of our theoretical analysis of improved resolution achieved by our method as compared to RP methods. We also present the necessary definitions to describe them. See Methods for details.

We define a discordant read as a read of a discordant pair and a concordant read as that of a concordant pair. Among the two reads of a pair, the one mapped upstream is called an upstream read and the other is called a downstream read in this paper. Let c be the depth of coverage. Assume that the positions of read pairs are uniformly random over the genome, and that the length r of each read is a fixed constant. Let q(c) be the probability that there is no read pair whose upstream read begins at a given base in the genome. Suppose that there are N read pairs uniquely mapped to a genomic sequence of length G. According to a classical analysis 18 ,

q ( c ) = 1 − 1 G N ≈ e − N / G = e − c / 2 r .

Hereafter, we just write q instead of q(c) for simplicity. In threshold-based RP approaches, the predicted position of an upstream end of a deletion is determined by the upstream discordant read that is the closest to the breakpoint. Let b be the position of an upstream end of a deletion, Δ _b be the distance between b and the closest upstream discordant read, and d be the distance between paired reads. We assume that d is a constant. Let E[Δ _b|b,c] be the expectation of Δ _bgiven that b is detected and the coverage is c. Then,

E [ Δ b | b , c ] = 1 − q 1 − q d + 1 S ( q , d ) ,

where

S ( q , d ) = ∑ j = 0 d j q j = q − ( d + 1 ) q d + 1 + d q d + 2 ( 1 − q ) 2 .

See Methods for derivation of Equation (2). We can obtain better resolution by using concordant reads in addition to discordant reads, because there is a chance that there exists a concordant read closer to b than any upstream discordant read (Figure 1). Such a read can contribute to the localization of the position where b can exist. Let Δ b ′ be the distance between b and the closest read in the upstream of b, and let E [ Δ b ′ | b , c ] be the expectation of Δ b ′ given that b is detected and the coverage is c. Then,

E [ Δ b ′ | b , c ] = 1 1 − q d + 1 × ( 1 − q 2 ) S ( q 2 , d ) − q d + 1 ( 1 − q ) S ( q , d ) .

Figure 1

Resolution improvement by exploiting concordant read pairs

Resolution improvement by exploiting concordant read pairs. Schematic illustration of the key idea of our method ChopSticks. Unlike conventional SV detection methods based only on discordant pairs whose mapping distances were not close to the expectation, ChopSticks uses concordant read pairs as well. There is a chance that there is a concordant read closer to the boundary of the deleted region (breakpoint) than any discordant reads. Such a concordant read localizes the predicted position of the breakpoint, and therefore it contributes to achieving a high resolution. In this figure, b is the upstream end of a true deletion, Δ_bis the distance between the upstream end of a true deletion and that of a deletion call by threshold-based read-pair (RP) methods. Similarly, Δ b ′ is defined for our method. The expected values of Δ_band Δ b ′ are given by Equations (2) and (3), respectively.

As shown in Figure 2, the expected resolution of our method is significantly superior to that of threshold-based RP methods, which only use discordant pairs. The achieved resolution is quite close to that of threshold-based RP methods but with double coverage, which we confirmed theoretically.

Figure 2

Expected resolutions of ChopSticks and threshold-based RP methods

Expected resolutions of ChopSticks and threshold-based RP methods. The expected resolution of our method (E [ Δ b ′ | b , c ] ) is shown by a thick red line, that of threshold-based RP methods (E[Δ_b|b,c]) is shown by a thin solid black line, and that of threshold-based RP methods with double coverage (E[Δ_b|b,2c]) is shown by a dashed black line. The difference between E [ Δ b ′ | b , c ] and E[Δ_b|b,2c] is also shown by a dotted blue line. As the coverage goes away from zero, the resolution obtained by our method quickly outperforms that of normal RP methods. It is also clear that the resolution of our method is very close to that of threshold-based RP methods with double coverage. The difference approaches zero when coverage approaches zero or infinity, as indicated by the blue dotted line. E[Δ_b|b,c], E [ Δ b ′ | b , c ] , and E[Δ_b|b,2c] are given by Equations (2), (3), and (5), respectively. In this figure, d=200 and r=100.

Theorem 1

The expectation E [ Δ b ′ | b , c ] is a weighted sum of E[Δ _b|b,2c] and E[Δ _b|b,c]. To be more precise, the following equation holds:

E [ Δ b ′ | b , c ] = ( 1 + q d + 1 ) E [ Δ b | b , 2 c ] − q d + 1 E [ Δ b | b , c ] .

See Methods for the proof. When c→0, both E[Δ _b|b,2c] and E[Δ _b|b,c] approach d/2, which is the expected resolution when a deletion is detected with only one read pair. Therefore E [ Δ b ′ | b , c ] also approaches d/2 when c→0. On the other hand, when c approaches infinity, E [ Δ b ′ | b , 2 c ] approaches E[Δ _b|b,2c] because q ^{d + 1}→0. In summary,

Theorem 2

E [ Δ b ′ | b , c ] is asymptotically equal to E[Δ _b|b,2c] when c→0 or c→∞.

Trimming of deletion calls to improve resolution

If all regions existing in the reference genome were covered by at least one read and there were absolutely no reads mapped to regions of homozygous deletions, the resolution of deletion calls could be quite easily improved by just trimming the ends of deletion calls that are covered by alignments of reads. Obviously, such a simple assumption does not hold in practical situations. First, coverage might be zero even in regions that actually exist in the genome, because no reads are obtained therein owing to the unevenness of the coverage or because reads cannot be uniquely mapped owing to repeat elements. Second, there might exist erroneous alignments in deleted regions because of incidental sequence similarity. Therefore, we developed the algorithm ChopSticks to carefully trim the ends of deletion calls (Figure 3). ChopSticks recognizes high-coverage regions close to the ends of deletion calls even if they are fragmented, and it repeatedly excludes the high-coverage regions from deletion calls. ChopSticks uses two parameters, k and f. The k parameter is a threshold used to distinguish high-coverage regions from low-coverage ones, and f determines the threshold of joint coverage of regions excluded from a deletion call. See Methods for details. Our implementation of ChopSticks is available on the Internet 19 .

Figure 3

Overview of trimming algorithm of ChopSticks

Overview of trimming algorithm of ChopSticks. Schematic illustration of the trimming algorithm of ChopSticks. ChopSticks trims ends of deletion calls that are not likely to be parts of deletions, according to their coverage. First, it trims high-coverage regions at the ends of deletion calls. Here, a high-coverage region is a region whose coverage is greater than a given parameter k. Second, it recognizes a high-coverage region separated by a low-coverage region and trims these regions if their joint coverage is deeper than kf, where f is another parameter. The second step is repeatedly conducted until the joint coverage becomes less than kf .

Computational experiment

To evaluate the power of ChopSticks in improving the resolution of deletion calls, we conducted computational experiments. Let the upstream difference of a deletion call be x−y, where x is the position of the upstream end of the true deletion and y be that of the deletion call. Similarly, let the downstream difference of a deletion call be y ^′−x ^′, where x ^′ is the position of the downstream end of the true deletion and y ^′is that of the deletion call. By definition, the closer to zero a difference is, the better. A positive difference value indicates that the called breakpoint is outside the true deletion, whereas a negative value indicates that it is inside the true deletion. To evaluate ChopSticks, the results of ChopSticks have to be compared with the positions of true deletions. Therefore we need NGS reads of a genome whose SVs against the reference genome are known up to bp-level resolution. We conducted two experiments described below.

Simulated reads

In the first experiment, we evaluated ChopSticks with simulated NGS reads for which all SVs were known up to bp-level resolution. To obtain data as realistic as possible, we generated a genome sequence with SVs and simulated NGS sequences by using SV annotations published by Quinlan et al. 7 . The accession number of the SV annotations is [dbVar:nstd19]. First, we deleted regions of the reference genome sequence that were annotated as deletions by Quinlan et al. Next, we inserted random fragments whose number and distribution of lengths were the same as annotated deletions, assuming that deletions and insertions are symmetric. Then, we introduced single nucleotide substitutions into the simulated genome sequence and generated paired reads from it. We conducted this simulation and evaluation of ChopSticks for chromosome 1 of the reference mouse genome mm9. All paired reads were mapped to mm9 using Burrows-Wheeler aligner (BWA) 20 . Then we conducted SV analysis by using SV detection tools from each of categories described in the Background section: BreakDancer 5 of threshold-based RP methods, MoDIL 8 of distribution-based RP methods, CLEVER 9 of graph-based RP methods, CNVnator 11 of RD methods, and Pindel 12 of SR methods. After that, we applied ChopSticks to their results.

Before applying ChopSticks, we examined the ability of SV detection tools to detect 460 deletions in chromosome 1 of the simulated mouse genome. We say that a deletion call is correct if it overlaps exactly one true deletion while the true deletion in turn overlaps exactly one deletion call. We show the number of called and correct SV calls in Table 1. We also show their recall (the number of correct deletion calls divided by the number of true deletions) and precision (the number of correct deletion calls divided by the number of all deletion calls) in Figure 4. The recall of BreakDancer and CLEVER was relatively good for all of tried coverage values, whereas the recall of Pindel was satisfactory only when coverage was high. The recall of MoDIL was low for all coverage values tried. Although almost all deletions called by these methods were correct, CNVnator generated numerous false positives (Table 1). Because ChopSticks is developed to correct breakpoints outside true deletions, we counted the number of deletion calls that cover the whole of true deletions. As shown in Figure 5, most of the deletion calls by MoDIL, CNVnator, and Pindel covered the whole of true deletions. However, a significant portion of BreakDancer and CLEVER results did not cover the whole of true deletions. Note that ChopSticks is harmless to these deletion calls because ChopSticks does not trim them when there are no alignments in true deletions.

Figure 4

Recall and precision of results of SV detection tools

Recall and precision of results of SV detection tools. BreakDancer and CLEVER achieved relatively good recall for all coverage, while recall of MoDIL was low. Although recall of CNVnator was not bad, its precision was low. The recall of an SR method Pindel was good when coverage was high, but it was insufficient when coverage was low.

Figure 5

Number of deletion calls covering the whole of true deletions

Number of deletion calls covering the whole of true deletions. Solid lines and circles show the number of all deletion calls generated by each tool, whereas dashed lines and ‘+’ symbol s show the number of deletion calls covering the whole of true deletions. Most of the deletion calls of MoDIL, CNVnator (expanded by the window size), and Pindel covered the whole of true deletions. On the other hand, many CLEVER results did not always contain the whole of true deletions, while median of the distribution of predicted breakpoints was close to the true breakpoints as shown in Figure 10. BreakDancer results for high coverage data did not always contain true deletions either. Predicted breakpoints of BreakDancer approached true breakpoints as the depth of coverage increases, and sometimes intruded into true deletions when coverage was high.

Table 1

Depth of coverage

SV caller

Results of deletion calls by BreakDancer, MoDIL, CLEVER, CNVnator, and Pindel. The values to the left of ”/” are the numbers of correct deletion calls, where a correct deletion call is the one that overlaps with exactly one true deletion, which, in turn, only overlaps with the deletion call; the values to the right of ”/” are the numbers of all deletion calls. BreakDancer and CLEVER results were good in both sensitivity and specificity. CNVnator generated numerous false positives, while Pindel suffered from low coverage. MoDIL missed lots of deletions.

BreakDancer

259/260

426/427

453/456

455/458

MoDIL

1/1

27/27

96/96

129/130

–/–

CLEVER

398/462

449/525

454/491

454/478

454/466

CNVnator

326/1,258

354/952

422/1,127

447/1,211

451/1,258

Pindel

85/85

317/317

436/438

450/454

456/456

Next, we applied ChopSticks to the results of SV detection tools. After that, we examined how well the resolution of deletion calls was improved. We tested ChopSticks for k=1,2,…,5 and f=0.1,0.2,…,1.0. We evaluated differences at both the upstream and downstream ends of deletions, and found that the results were similar. Therefore we only present the results at upstream ends.

Resolution improvements for BreakDancer deletion calls

As shown in Figure 6, the resolution of deletion calls was clearly improved by using ChopSticks. The original BreakDancer results was successfully corrected, which is also clear in Figure 7. When coverage was low, the resolution was well improved for small k values. When coverage was high, the resolution was also improved for large k values. Therefore, when the coverage is high, we recommend using large k values to ignore erroneous alignments. As shown in Figure 8, ChopSticks worked well regardless of deletion lengths.

Figure 6

BreakDancer results improved by ChopSticks

BreakDancer results improved by ChopSticks. Box-and-whisker plots of upstream differences of deletion calls obtained by BreakDancer and those improved by ChopSticks. The red, green, blue, light blue, and magenta boxes correspond to k values of 1, 2, 3, 4, and 5, respectively, and the rightmost yellow box corresponds to the original results of BreakDancer. Among boxes of the same color, from left to right, f=0.1, 0.2, …, 1.0. Brown horizontal dashed lines indicate the values of 25%, 50%, and 75% tiles of differences of original deletion calls from below to above, respectively. The results in this figure indicate that ChopSticks clearly improved the resolution of the original BreakDancer results. When the coverage was low, small k values were effective in improving the resolution. When coverage was high, the resolution was also improved for large k values. Therefore, when the coverage is high, we recommend using large k values to avoid erroneous alignments of NGS reads and the genome. We omitted the results for coverage=15 because they were similar to those for coverage=20.

Figure 7

Distribution of differences of BreakDancer results and those improved by ChopSticks

Distribution of differences of BreakDancer results and those improved by ChopSticks. The distribution of differences of ChopSticks results concentrated around zero, whereas that of BreakDancer results had long tail in 0–50 bp. Here, k=2, f=0.5, and coverage=5. Each frequency corresponds to the number of differences in bins of 2 bp.

Figure 8

Scatter plot of deletion lengths and differences of deletion calls

Scatter plot of deletion lengths and differences of deletion calls. No correlation between deletion lengths and differences was observed (r²=0.056). ChopSticks worked well regardless of deletion lengths. Here, k=2, f=0.5, and coverage=5.

Resolution improvements for MoDIL deletion calls

As shown in Figure 9, the resolution of deletion calls by MoDIL was also improved by using ChopSticks. We omitted evaluation of MoDIL for coverage=20 because MoDIL was very slow (See Methods).

Figure 9

MoDIL results improved by ChopSticks

MoDIL results improved by ChopSticks. Box-and-whisker plots of upstream differences of deletion calls obtained by MoDIL and those improved by ChopSticks. The format of this plot is exactly the same as that in Figure 6, except that results for coverage=15 were shown instead of those for coverage=20. The results in this figure indicate that ChopSticks can also improve the resolution of MoDIL results.

Resolution improvements for CLEVER deletion calls

The resolution of deletion calls by CLEVER was also improved by using ChopSticks. As mentioned above, deletion calls of CLEVER do not always cover the whole of true deletions. Nonetheless, as shown in Figure 10 and 11, ChopSticks successfully improved resolution of CLEVER results by selectively correcting predicted breakpoints outside true deletions.

Figure 10

CLEVER results improved by ChopSticks

CLEVER results improved by ChopSticks. Box-and-whisker plots of upstream differences of deletion calls obtained by CLEVER and those improved by ChopSticks. The differences were successfully corrected. Note that a significant portion of breakpoints predicted by CLEVER were inside the true deletion. Nonetheless, ChopSticks selectively trimmed predicted breakpoints outside true deletions, and left those inside untouched.

Figure 11

Distribution of differences of CLEVER results and those improved by ChopSticks

Distribution of differences of CLEVER results and those improved by ChopSticks. The distribution of differences of CLEVER results had long tail in 0–50 bp, whereas that improved by ChopSticks concentrates around zero. Here, k=2, f=0.5, and coverage=5. Each frequency corresponds to the number of displacements in bins of 2 bp.

Resolution improvements for CNVnator deletion calls

Because RD methods call SVs by examining coverages in windows of a fixed size, the positions of breakpoints predicted by the RD methods have unavoidable ambiguity and they might be either inside or outside true deletions. Because ChopSticks assumes that predicted breakpoints are outside true deletions, we applied ChopSticks after we expanded deletion calls of CNVnator at both ends by the window size. As shown in Figure 12, the results of CNVnator were successfully improved. This result indicates that ChopSticks is also available for RD methods in addition to RP methods.

Figure 12

CNVnator results improved by ChopSticks

CNVnator results improved by ChopSticks. Box-and-whisker plots of upstream differences of deletion calls obtained by CNVnator and those improved by ChopSticks. The format of this plot is exactly the same as that in Figure 6. We expanded the original deletion calls of CNVnator outward by the window size (50 bp) because ChopSticks assumes that predicted breakpoints are outside true deletions. The results in this figure indicate that ChopSticks can improve the resolution of CNVnator results if predicted positions of breakpoints are within a few hundreds of bases from true breakpoints.

Results of ChopSticks applied to Pindel deletion calls

Owing to the SR signature that allows Pindel to detect SVs at bp-level resolution, the positions of breakpoints obtained with Pindel were quite accurate. When ChopSticks was applied to the results of Pindel, the results became slightly worse than the original Pindel results, as shown in Figure 13, although differences remained close to zero in most cases. Note that the recall of Pindel was not satisfactory when coverage is low, as shown in Figure 4. ChopSticks is useful in cases where deletions missed by Pindel are analyzed.

Figure 13

Pindel results and those modified by ChopSticks

Pindel results and those modified by ChopSticks. Box-and-whisker plots of upstream differences of deletion calls obtained by Pindel and those modified by ChopSticks. The format of this plot is exactly the same as in Figure 6. The results in this figure indicate that ChopSticks should not be applied to the Pindel results because the resolution of the Pindel results is already quite high.

Real Illumina reads of DBA/2J

In the second experiment, we evaluated ChopSticks using the real NGS sequences of Quinlan et al. 7 . The sample was taken from a female mouse of the DBA/2J strain, whose genome contains SVs against the reference genome of the C57BL/6J strain 21 . The read sequences were available from the NCBI Sequence Read Archive (SRA) database 22 . The accession number of the read sequences is [SRA:SRA010027]. To evaluate the results of ChopSticks, we need bp-level SV annotations of DBA/2J as well. Therefore we generated deletion calls at bp-level resolution using Sanger reads in a manner similar to that of Quinlan et al. See Methods for details. Our deletion calls are available at the dbVar database under accession no. [dbVar:nstd70].

We tried the five SV detection tools used in the previous experiment, and found that MoDIL, CNVnator and Pindel missed the most of deletions detected with Sanger reads. These methods seemed to suffer from the low depth of coverage and short read lengths. Therefore, we hereafter only describe results of ChopSticks applied to BreakDancer and CLEVER results.

Resolution improvements for BreakDancer deletion calls

Figure 14 shows the differences between BreakDancer results and those improved by using ChopSticks. As the previous experiment where simulated NGS reads were used, the differences obtained with real NGS reads were reduced. The median and differences less than the median clearly shifted toward zero, which is also clear in Figure 15. Although ChopSticks trimmed some deletion calls into those based on Sanger reads when k=1 or k=2 and f was small, this problem quickly disappeared as k or f became larger. No correlation between deletion lengths and the performance of ChopSticks were observed (r ²=0.021). Although we generated 525 deletion calls by using Sanger reads, only 83 of them were found by BreakDancer. There were at least two reasons for this difference in numbers. First, it is difficult to find small deletions because read pairs spanning small deletions might not be recognized as discordant pairs. Second, a lot of deletion calls based on Sanger reads had fewer than two NGS-read pairs spanning them. Such deletion calls would be missed because BreakDancer deletion calls must be supported by at least two pairs when the default parameters are used, in order to reduce false positives. For this data set, 82 of all 83 deletion calls generated by BreakDancer contained the whole of deletions predicted with Sanger reads.

Figure 14

BreakDancer results for DBA/2J reads improved by ChopSticks

BreakDancer results for DBA/2J reads improved by ChopSticks. Box-and-whisker plots of upstream and downstream differences of deletion calls obtained by BreakDancer and those improved by ChopSticks. The results in this figure indicate that ChopSticks can improve the resolution of deletion calls for real sequences. Although ChopSticks trimmed upstream ends of a few deletion calls too much when k=1 or k=2 and f was small, such problems quickly disappeared for greater k and f values.

Figure 15

Distribution of differences of BreakDancer results and those improved by ChopSticks.

Distribution of differences of BreakDancer results and those improved by ChopSticks. The distribution of differences of BreakDancer results had long tail in 0–400 bp, whereas that improved by ChopSticks concentrates around zero and frequencies in the long tail were reduced. Here, k=2, f=0.5. Each frequency corresponds to the number of differences in bins of 20 bp.

Resolution improvements for CLEVER deletion calls

CLEVER detected much more (347) deletions than BreakDancer. The results of CLEVER were also improved by ChopSticks as shown in Figure 16, where the peak around zero became stronger. However, it was difficult for ChopSticks to correct positions of breakpoints when they were away from those predicted with Sanger reads by hundreds of bases.

Figure 16

Distribution of differences of CLEVER results and those improved by ChopSticks

Distribution of differences of CLEVER results and those improved by ChopSticks. ChopSticks corrected some of breakpoints predicted by CLEVER so that the peak at zero became stronger. However, the distribution of differences of CLEVER results had long tail in 0–3000 bp and it was difficult for ChopSticks to correct such large differences. Here, k=2, f=0.5. Each frequency corresponds to the number of differences in bins of 20 bp.

Conclusions

We have presented a new method called ChopSticks to improve the resolution of predicted positions of deletions. The key idea is to exploit both concordant read pairs and discordant ones. According to our theoretical analysis, the resolution of our method is quite similar to that of threshold-based RP methods but with double coverage. In an experiment on simulated NGS reads, ChopSticks clearly improved the results of BreakDancer, MoDIL, CLEVER, and CNVnator. Although the resolution of Pindel results is quite high, ChopSticks works well even for low-coverage data where recall of Pindel is not sufficient. The effectiveness of ChopSticks was also confirmed by performing an experiment on real Illumina reads. Despite a number of methods proposed for detecting SVs 2 3 4 , there is no one-stop method that simultaneously achieves high sensitivity, high specificity, high resolution, and robustness for low-coverage data. Therefore a combination of SV detection methods is required, and ChopSticks can play an important role because it uses new independent information ignored in other methods.

As a future work, we consider to develop a method to distinguish homozygous deletions from heterozygous ones and to apply ChopSticks to the former. With this approach, ChopSticks will be available for more applications.

Methods

Derivation of theoretical estimation of resolution

Because the resolution at downstream ends of deletions can be estimated symmetrically, we only analyze the resolution at upstream ends. Let P _bbe the probability that a breakpoint b is successfully included in a deletion call by a threshold-based RP method. If b is detected, there exists an upstream discordant read within d bases from b. Therefore,

P b = 1 − q d + 1 .

We derive the expected distance between the true ends of deletions and the predicted ones in a manner similar to Bashir’s analysis 23 . For 0≤j≤d, Bashir et al. defined A _jas an event in which b is detected and an upstream read of a discordant pair is exactly j bases upstream of b. The probability that A _joccurs is

Pr ( A j ) = ( 1 − q ) q j .

Consequently,

E [ Δ b | b , c ] = 1 P b ∑ 0 ≤ j ≤ d j Pr ( A j ) = 1 − q 1 − q d + 1 S ( q , d ) .

Similarly, we define A j ′ as an event wherein b is detected and the closest read upstream of b is exactly j bases apart. There are two mutually exclusive cases: (i) at least one of the closest reads is an upstream discordant read or (ii) all the closest reads are concordant reads. In the latter case, we have to consider the joint probability of the following events.

• A concordant read exists at j bases upstream of b, the probability of which is 1−q.

• No read nearer than the closest concordant read exists, the probability of which is q ^2j.

• No discordant read exists at j bases upstream of b, the probability of which is q.

• There must exist an upstream read of discordant pairs whose alignment ends in a region that is j + 1 to d bases upstream of b so that b is successfully included in a deletion call, the probability of which is 1−q ^d−j.

Therefore,

Pr ( A j ′ ) = ( 1 − q ) q 2 j + ( 1 − q ) q 2 j q ( 1 − q d − j ) = ( 1 − q 2 ) q 2 j − q d + 1 ( 1 − q ) q j .

Consequently,

E [ Δ b ′ | b , c ] = 1 P b ∑ 0 ≤ j ≤ d j Pr ( A j ′ ) = 1 1 − q d + 1 ( 1 − q 2 ) S ( q 2 , d ) − q d + 1 ( 1 − q ) S ( q , d ) .

Proof of Theorem 1

From Equation (1), E[Δ _b|b,2c] can be obtained by replacing qwith q ² in Equation (2):

E [ Δ b | b , 2 c ] = 1 − q 2 1 − q 2 ( d + 1 ) S ( q 2 , d ) .

From Equations (2), (3), and (5), Equation (4) can be obtained.

Proof of Theorem 2

First, we consider a case where c→0. Because q→1 by Equation (1),

S ( q , d ) → ∑ j = 0 d j = d ( d + 1 ) 2 .

Besides,

1 − q 1 − q d + 1 = 1 1 + q + q 2 + ⋯ + q d → 1 d + 1 .

Therefore, all of E [ Δ b ′ | b , c ] , E[Δ _b|b,2c], and E[Δ _b|b,c] approach d/2 by Equation (4). On the other hand, when c→∞, q ^{d + 1}approaches 0. In consequence, the right hand side of Equation (4) approaches E[Δ _b|b,2c] when c→0 or c→∞.

Mapping to the genome

We mapped paired reads to the mm9 reference genome sequences of Mus musculus using BWA version 0.5.9 20 with default parameters. The target genome sequences involved in our experiment included all chromosomes of mm9 except chromosome Y, assuming cases where a female mouse was analyzed 7 21 .

Simulated NGS sequences

To focus on uniquely mapped reads for BreakDancer, MoDIL, CLEVER, and ChopSticks, we removed paired reads if the mapping quality (MAPQ) score was zero for at least one of the two reads of a pair. For CNVnator and Pindel, we used the result of BWA without filtering. We show the total length of reads and the number of aligned reads in Table 2.

Table 2

Depth of coverage

Summarized statistics of simulated NGS reads and their alignments to mm9. On the third row, we only counted read pairs whose reads were both mapped uniquely.

Total number of bases

394,391,200

985,978,000

1,971,956,000

2,957,934,000

3,943,912,000

Number of reads

3,943,912

9,859,780

19,719,560

29,579,340

39,439,120

Number of mapped reads

3,677,398

9,194,942

18,391,288

27,587,970

36,783,348

Real DBA/2J sequences

We split the data set of NGS reads into 275 subsets, and mapped each of them with an independent BWA process and merged the results. Then we removed reads whose MAPQ score was zero for at least one of the two reads of a pair. We show the total length of reads and the number of aligned reads in Table 3.

Table 3

Summarized statistics of NGS reads of the DBA/2J strain 7 and their alignments to mm9.

Total number of bases

13,050,980,662

Number of reads

330,462,408

Reads of uniquely mapped pairs

149,021,716

Reads of uniquely mapped pairs (chromosome 1)

10,316,525

Trimming algorithm of ChopSticks

The coverage outside a deletion should be higher than that inside it. Therefore ChopSticks repeatedly recognizes a high-coverage region in a deletion call that is likely a continuation of a high-coverage region outside the deletion. We show in Figure 17 the trimming algorithm executed by ChopSticks for upstream ends. Here is a brief description of the algorithm: Line 2: Skip a high-coverage region at the end of the deletion call. Lines 6–9: Go through a low-coverage region. Lines 10–13: Go through a high-coverage region. Line 14: If the joint coverage is low, exit the loop. Line 17: Trim regions which the algorithm has gone through.Trimming of the downstream ends is conducted symmetrically.

Figure 17

Pseudocode of trimming algorithm

Pseudocode of trimming algorithm. Pseudocode of the trimming algorithm of ChopSticks. Here, L is the length of the deletion call being processed, k is a threshold used to discriminate high-coverage regions from low-coverage ones, and f is a parameter that determines the threshold of the coverage of regions to be trimmed. The variable x represents the position of the base being examined, and the variable y represents the length of a region to be trimmed. The value c[x] is the coverage at the x-th base in the deletion call, while s keeps the sum of c[x] values.

Data for computational experiments

To evaluate our method, we need NGS sequences and reliable bp-level positions of breakpoints. There were six SV studies of inbred mice (nstd5, 7, 15, 18, 19, and 48) in the dbVar database 22 when we accessed it on April 1, 2012. However, none of them provides accurate bp-level positions of breakpoints. Therefore, we evaluated ChopSticks using the following two data sets.

Simulated NGS reads

We artificially introduced deletions and insertions into the mm9 reference genome and then generated simulated NGS reads using the modified genome. To obtain most realistic simulated sequences, we built a simulated genome sequence using SV annotations generated by Quinlan et al. 7 , which are available from the dbVar database under accession no. [dbVar:nstd19]. First, we deleted regions annotated as deletions in [dbVar:nstd19] from the mm9 reference genome sequence of chromosome 1. We show the distribution of lengths of deletions in Figure 18. Second, we inserted fragments consisting of randomly chosen bases so that the number and the distribution of lengths of inserted fragments were the same as those of deletions, assuming that the genome to be analyzed and the reference genome are affected symmetrically by deletions and insertions. Third, we introduced random single nucleotide substitutions with a probability of 1.0×10⁻⁴ at each base. Finally, we generated paired reads from the modified genome sequence so that the read length was 100 bp and the average and the standard deviation of distances of paired reads were 200 bp and 50 bp, respectively. We generated five sets of simulated NGS reads whose depth of coverage were 2, 5, 10, 15, and 20, respectively.

Figure 18

Distribution of deletion lengths in our simulation

Distribution of deletion lengths in our simulation.

NGS reads of Quinlan et al. and deletion calls based on Sanger reads

We generated our own bp-level deletion calls by using publicly available Sanger reads of the DBA/2J strain. From the NCBI trace archive, we retrieved all 7,998,826 Sanger reads of whole-genome shotgun sequencing for the DBA/2J strain. We mapped these Sanger reads to chromosome 1 of mm9 by MegaBLAST 24 , and we searched for Sanger reads that were split into two parts and aligned uniquely on the same strand and in the right order. There were 763 reads that indicated deletions whose lengths were at least 50 bp. By merging redundant ones, we obtained 525 deletion calls. These deletion calls are available in the dbVar database under accession no. [dbVar:nstd70]. We show the distribution of their lengths in Figure 19. NGS sequences of the DBA/2J strain generated by Quinlan et al. are available in the SRA database 22 under accession no. [SRA:SRA010027].

Figure 19

Distribution of deletion lengths detected with Sanger reads

Distribution of deletion lengths detected with Sanger reads.

Parameters for SV detection tools and evaluation of their results

We executed BreakDancer with default parameters, and Pindel with an expected template size of 432 bp because the median fragment size was 432 bp according to Quinlan et al. 7 . For CNVnator, we tested three window sizes: 50 bp, 100 bp, and 200 bp. Because the recall of window size 50 bp outperformed those of window sizes 100 bp and 200 bp for our simulated data when coverage was 2, we used results of window size 50 bp for evaluation. Because CLEVER tends to generate deletion calls duplicatedly with slightly different positions, we chose the best one for those overlapping with true deletions in order to estimate the upper limit of the accuracy of CLEVER. We divided the chromosome 1 of mm9 into 5.1 Mbp fragments in a manner such that franking fragments share 0.1Mbp, and applied MoDIL to each fragments, because MoDIL was quite slow as reported previously 9 . We omitted evaluation of MoDIL for coverage=20.

To compare the positions of true and predicted deletions, we used BEDTools 25 .

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

TY conceived the project, invented and implemented the algorithms, and performed the computational analysis. SS assisted TY in conducting experiments. MN and SM critically revised the manuscript. All authors read and approved the final manuscript

Acknowledgements

The super-computing resource was provided by Human Genome Center (University of Tokyo).

Illumina Sequencing portfolio[http://www.illumina.com/systems/sequencing.ilmn]A map of human genome variation from population-scale sequencingThe 1000 genomes project consortiumNature20104671061107310.1038/nature095343042601209810921000 genomes project: Mapping copy number variation by population-scale genome sequencingMillsREWalterKStewartCHandsakerREChenKAlkanCAbyzovAYoonSCYeKCheethamRKChinwallaAConradDFFuYGrubertFHajirasoulihaIHormozdiariFIakouchevaLMIqbalZKangSKiddJMKonkelMKKornJKhuranaEKuralDLamHYKLengJLiRLiYLinCYLuoRNature2011470596510.1038/nature09708307705021293372Computational methods for discovering structural variation with next-generation sequencingMedvedevPStanciuMBrudnoMNat Methods20096S13—S2019844226BreakDancer: an algorithm for high-resolution mapping of genomic structural variationChenKWallisJWMcLellanMDLarsonDEKalickiJMPohlCSMcGrathSDWendlMCZhangQLockeDPShiXFultonRSLeyTJWilsonRKDingLMardisERNat Methods2009667768110.1038/nmeth.136319668202Combinatorial algorithms for structural variation detection in high-throughput sequenced genomesHormozdiariFAlkanCEichlerEESahinalpSCGenome Res2009191270127810.1101/gr.088633.108270442919447966Genome-wide mapping and assembly of structural variant breakpoints in the mouse genomeQuinlanARClarkRASokolovaSLeibowitzMLZhangYHurlesMEMellJCHallIMGenome Res20102062363510.1101/gr.102970.109286016420308636MoDIL: detecting small indels from clone-end sequencing with mixtures of distributionsLeeSHormozdiariFAlkanCBrudnoMNat Methods2009647347410.1038/nmeth.f.25619483690CLEVER: Clique-Enumerating Variant FinderMarschallTCostaICanzarSBauerMKlauGSchliepASchönhuth ABioinformatics2012,282875288210.1093/bioinformatics/bts56623060616Identification of somatically acquired rearrangements in cancer using genome-wide massively parallel paired-end sequencingCampbellPJStephensPJPleasanceEDO’MearaSLiHSantariusTStebbingsLALeroyCEdkinsSHardyCTeagueJWMenziesAGoodheadITurnerDJCleeCMQuailMACoxABrownCDurbinRHurlesMEBignellPAWEGRandP Andrew FutrealMRSNat Genet20084072272910.1038/ng.128270583818438408CNVnator: an approach to discover, genotype, and characterize typical and atypical CNVs from family and population genome sequencingAbyzovAUrbanAESnyderMGersteinMGenome Res20112197498410.1101/gr.114876.110310633021324876Pindel: a pattern growth approach to detect break points of large deletions and medium sized insertions from paired-end short readsYeKSchulzMHLongQApweilerRNingZBioinformatics200925212865287110.1093/bioinformatics/btp394278175019561018De novo assembly of human genomes with massively parallel short read sequencingLiRZhuHRuanJQianWFangXShiZLiYLiSShanGKristiansenKLiSYangHWangJWangJGenome Res20102026527210.1101/gr.097261.109281348220019144Target-enrichment strategies for next-generation sequencingMamanovaLCoffeyAJScottCEKozarewaITurnerEHKumarAHowardEShendureJTurnerDJNat Methods2010711111810.1038/nmeth.141920111037Discovery and genotyping of genome structural polymorphism by sequencing on a population scaleHandsakerRKornJNemeshJMcCarrollSNat Genet201143326927610.1038/ng.76821317889SVseq: an approach for detecting exact breakpoints of deletions with low-coverage sequence dataZhangJWuYBioinformatics201127233228323410.1093/bioinformatics/btr56321994222Inbred mice - The Jaxon Laboratory[http://jaxmice.jax.org/type/inbred/index.html]Genomic mapping by fingerprinting random clones: a mathematical analysisLanderESWatermanMSGenomics19882323123910.1016/0888-7543(88)90007-93294162ChopSticks[https://github.com/toyasuda/ChopSticks]Fast and accurate short read alignment with Burrows-Wheeler transformLiRZhuHRuanJQianWFangXShiZLiYLiSShanGKristiansenKLiSYangHWangJWangJBioinformatics2009251754176010.1093/bioinformatics/btp324270523419451168Initial sequencing and comparative analysis of the mouse genomeMouse genome sequencing consortiumNature200242052056210.1038/nature0126212466850Database resources of the National Center for Biotechnology InformationSayersEWBarrettTBensonDABoltonEBryantSHCaneseKChetverninVChurchDMDiCuccioMFederhenSFeoloMFingermanIMGeerLYHelmbergWKapustinYLandsmanDLipmanDJLuZMaddenTLMadejTMaglottDRMarchler-BauerAMillerVMizrachiIOstellJPanchenkoAPhanLPruittKDSchulerGDSequeira ENuc Acids Res201139suppl 1D38—D51Evaluation of paired-end sequencing strategies for detection of genome rearrangements in cancerBashirAVolikSCollinsCBafnaVRaphaelBJPLoS Comput Biol20084e100005110.1371/journal.pcbi.1000051227837518404202A greedy algorithm for aligning DNA sequencesZhangZSchwartzSWagnerLMillerWJ Comp Biol2000720321410.1089/10665270050081478BEDTools: a flexible suite of utilities for comparing genomic featuresQuinlanARHallIMBioinformatics201026684184210.1093/bioinformatics/btq033283282420110278