Overlapping paired-end genome sequencing of polyhomeotic tumors

To assess the relevance of genome integrity in Drosophila tumorigenesis we took advantage of a tumor model induced by downregulation of the tumor suppressor gene polyhomeotic in the posterior compartment of the wing disc [25], [26]. By using the GAL4-UAS system [27] a spatially and temporally controlled knockdown of tumor suppressor genes can be achieved. In our model GAL4 is specifically expressed within the posterior compartment of the wing imaginal disc (see Materials and Methods), leading to a posterior-confined expression of an RFP gene (UAS-RFP) and an interfering double stranded RNA targeting ph (UAS-phRNAi). Figure 1A and B depict examples of both a wild type disc and a tumor, respectively. The penetrance of the phRNAi tumor phenotype, indicating the probability of tumor development, was approximately 0.15 (data not shown). phRNAi tumors exhibit extensive morphological anomalies (Figure 1B and Figure S1A) and are characterized by fast growth, high cell density, and a loss of tissue polarity. Another transgene present in our tumor model is a Notch signaling pathway reporter expressing EGFP in response to Notch activity (NRE:EGFP). In wild-type discs, high Notch activity is restricted to the dorsal-ventral boundary (Figure 1A). However, upon Ph depletion we observed ectopic activation of the Notch signaling pathway (Figure 1B), which contributes to tumor growth [26]. The anterior compartment of the tumor, not affected by the RNAi, is clearly distinguishable from the tumorigenic tissue and Notch activity in this compartment resembles the normal pattern observed within the control (Figure 1C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Depletion of polyhomeotic (ph) induces neoplastic tumors.

A) Ph expression in the normal wing disc of third instar larvae (left) expressing the reporter en-GAL4 UAS-myr-RFP, UAS-Dicer2, NRE:EGFP (right). B) Downregulation of Ph induced by the RNAi reporter observed in the posterior compartment. Posterior compartment (red RFP) shows overproliferation phenotype (from [26]). C) Schematic outline of the experimental workflow. Eggs of the same genotype were developed at different temperatures. The tumor suppressor ph is specifically depleted (RNAi) at 25°C within the posterior compartment (p, RFP signal in red) of wing imaginal discs, leading to the formation of large tumors (upper). To allow for the accumulation of SVs, tumors are transplanted for a period of four weeks. At 18°C depletion of Ph is not sufficient to drive tumorigenesis (lower) and corresponding wing imaginal discs were used as control. Genomic DNA from both samples was isolated and subjected to paired-end sequencing. Notch-dependent EGFP expression (green) marks the boundary of the dorsal and ventral compartments. The white dashed outline marks the remnant anterior compartment (a) with normal Notch signaling along the dorsal/ventral boundary, while the grey dashed outline labels the haltere disc (hd).

https://doi.org/10.1371/journal.pone.0087090.g001

The acquisition of genomic aberrations during tumorigenesis is thought to be a gradual process, increasing in frequency as the cellular DNA damage response mechanisms become progressively affected by the onset of the disease. To increase the incidence probability of genomic aberrations in our tumor model, we prolonged the tumor growth period by repeated transplantation and culturing of allografts of the tumorigenic tissue in the abdomen of host flies (Figure S1B), over a total period of four weeks. A schematic representation of our experimental workflow is illustrated in Figure 1C. Notably, we observed a comparable growth characteristic and morphology in allografts of tumorigenic wing tissue from genetic ph knockout mutants (Figure S1C) ruling out tumor promoting capacities by secondary RNAi effects or by the stimulation of the RNAi machinery.

Next, we collected control material of the same genotype and from the same tissue but without neoplastic characteristics, to be able to distinguish somatic mutations and aberrations from germline events with high accuracy. For this purpose, we took advantage of the temperature sensitivity of the GAL4-UAS system. RNAi in our transgenic Drosophila model is temperature- and dosage-sensitive allowing for a broad range of allelic series (full knockdown at vs. partial knockdown at ) [26] to be generated. Accordingly, while depletion of Ph at was insufficient to induce overproliferation, offspring larvae of the very same mothers kept at developed highly proliferative neoplastic tumors (Figure 1C). Thus, genomic DNA was isolated from control and tumor wing discs, developed at and , respectively. Next, we generated genomic DNA libraries exhibiting mean fragment sizes of approximately 250 bp (Figure S1D–E) and subjected these fragments to 150 bp overlapping paired-end sequencing, resulting in read pairs featuring a central overlap. Table 1 summarizes the sequencing results used for the subsequent analysis of the genome structure.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of the sequencing experiments and the seed-based alignment classification.

https://doi.org/10.1371/journal.pone.0087090.t001

Since cancer progression has been associated with gross chromosomal rearrangements and aneuploidy, leading to copy number variations (CNV) of large genomic regions and karyotype changes [10], we used BICSeq, a CNV detection algorithm [28], to disclose such events in our samples. However, a global comparison of the tumor and the control coverage did not indicate the presence of large CNVs (Figure S2A). Another feature of cancer genomes, recently discovered in genome-wide studies of breast cancer, are mutational processes leaving specific signatures of base substitutions [3]. To investigate whether a similar process is activated during neoplastic growth in our tumor model, we analyzed the prevalence of single base substitutions [29] in the tumor versus control, but similarly, no substantial differences could be detected (Figure S2B). Therefore, we proceeded further and searched for tumor specific SVs at higher resolution using the DSVD algorithm outlined below.

Overlapping read pairs allow exact alignment classification

The resolution of paired-end read-based strategies for SV detection is limited by the uncertainty within the fragment length, reflected by the fragment size distribution. However, advancements in NGS technology enable the generation of sufficiently long read pairs such that each read covers the central region of the fragment, while maintaining the fragment size distribution approximately constant. Accordingly, the centrally overlapping parts of read pairs can be used for the exact reconstruction of the fragment sequence [21]. Figure S3A shows a schematic representation of the procedure. As outlined below, the fragment sequence allows for both an exact evaluation of the mapping distance () of the read pair alignment and a precise identification of the SV breakpoints. The detection and characterization of SVs requires the alignment of read pairs to a reference genome. As most informative fragments are derived from regions exhibiting systematic differences with respect to the reference genome, the presence of SVs is likely to impair the alignment and identification of discordant read pairs, ultimately leading to a loss of relevant information (see Figure 2A). To overcome this problem, we performed a seed-based alignment (Figure 2A) using seeds, subsequences derived from the high-confidence 5′-ends of the reads, similar to [30]. The choice of the seed length constitutes a tradeoff between ‘unique alignability’, defined as the probability that a randomly selected chromosomal subsequence of length can be uniquely aligned to its origin, and recall, i.e. the ability to align the seed at all. Based on the evaluation of the unique alignability we determined a seed length to be eligible for the Drosophila reference genome (Figure S3B–D). In order to account for SNPs and sequencing errors, seeds were aligned allowing one mismatch [31]. Furthermore, we required that both seeds of a read pair uniquely align to exactly one position in the genome. To identify discordant alignments the length of the afore reconstructed fragment sequence is employed. Based on the mapping distance of the seed alignment the following classification rule is defined: if the alignment and hence the read pair is considered to be concordant, if the alignment and hence the read pair is classified as discordant. In addition, the relative orientation of the seeds is taken into account during the classification. Consequently, by considering the fragment sequence and the seed alignments, discordant read pairs are identified at single base resolution. In addition to the single base-resolved classification of the read pairs, we employed the reconstructed fragment sequences for a precise characterization of SVs. Genomic rearrangements result in aberrant DNA sequences joining different parts of the genome (Figure 2A). The exact identification of breakpoints requires the mapping of the different fragment subsequences to their distinct origins within the reference genome. Based on the discordant seed alignments, we constructed a ‘minimal reference’ (Figure 2B, left), confining the search space to the smallest conceivable region of the reference genome possibly containing the fragment sequence. Subsequently, a pairwise global alignment involving the minimal reference and the reconstructed fragment sequence is performed resulting in a gapped alignment (see Materials and Methods). The genomic coordinates of the breakpoints can be inferred from start and end position of the gap (Figure 2B, right).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Schematic representation of the SV detection performed by DSVD.

(A) Genomic rearrangements join different regions of the genome resulting in aberrant sequences. As a consequence, the full-length read alignment may fail. To avoid this problem, seeds of length , derived from the 5′-ends of the reads, are aligned instead. (B) Left: first, we constructed a minimal reference, i.e. the smallest possible region of the reference sequence possibly containing the reconstructed fragment. The construction requires the extension of the reference sequence in correct orientation, starting at the seed alignments, to a total length equal to where corresponds to length of the fragment sequence. The two extensions are ultimately joined to form the minimal reference. Right: next, for an exact identification of the breakpoints a global alignment of the reconstructed fragment sequence and the minimal reference is performed using the Needleman-Wunsch algorithm (see Materials and Methods). The gapped alignment is subsequently used to identify the breakpoint positions corresponding to the start and end position of the gap. (C) Schematic representation of an insertional duplication with inversion of the upstream inserted sequence (blue). Sequencing and processing of the two fragments (frag1 and frag2 in orange and red, respectively) spanning the boundaries of the aberrant region lead to the identification of the insertion site (bp 1) and the two virtual breakpoints (bp 2 and bp 3). The dashed lines connecting the reads of a pair derived from the fragments establish particular connections (Table 3) between different breakpoints resulting in an SV-type specific signature on the reference genome. In this example frag1 connects bp 1 and bp 3, approaching both breakpoints from the left. Similarly, frag2 establishes a connection between bp 1 and bp 2, approaching either genomic coordinate from the right. The connections formed by the read pairs can be represented explicitly by introducing directed edges between the different breakpoints. (D) Upper: schematic representation of the discordant graph representing all identified SVs. The example in the inset is similar to C. Lower: the prototype graph for the SV outlined in C. The graph structure represents the signature resulting from the duplication and inversion of the region between bp 2 and bp 3 followed by an upstream insertion at bp 1. Dashed lines highlight an isomorphism between the vertices of the prototype graph and the SV-representing subgraph, since the existence of an edge between two vertices in implies the existence of the same edge connecting the transformed vertices in .

https://doi.org/10.1371/journal.pone.0087090.g002

Graph-based representation of structural variations

The applicability of different classes of graphs (e.g. undirected or de Brujin graphs) for a comprehensive representation and identification of SVs appears natural in this context and was successfully demonstrated before [32], [33]. In our approach, we integrate information represented by discordant read pairs, which relate different breakpoints, by using directed weighted multigraphs, thus, achieving a comprehensive representation of a broad range of SVs (see Materials and Methods). An exemplary SV is illustrated in Figure 2C.

To summarize and relate all breakpoints detected by all the discordant read pairs, we construct the so-called ‘discordant graph’ (Figure 2D, upper), which represents the entire information reflecting structural differences as detected by the sequencing experiment. In the discordant graph, vertices represent genomic coordinates of the breakpoints, edges correspond to connections established by read pairs (see enlargement in Figure 2D) and the weight of an edge is equal to the number of read pairs supporting this connection. The discordant graph is disconnected and consists of maximally connected subgraphs (connected components) corresponding to the different SVs (Figure 2D, upper).

The correct interpretation of these components is fundamental for the identification of the SV types they represent. The inset in Figure 2D illustrates a subgraph consisting of three breakpoints connected by edges of different types. This example shows that type and characteristics of the corresponding SV are difficult to infer by visual inspection. To overcome this problem, we defined so-called ‘prototype graphs’ and use them to search the discordant graph. For example, the prototype corresponding to an upstream insertional duplication with inversion (Figure 1C) is illustrated in Figure 2D (lower). The problem of identifying SVs can now be stated as follows: given a prototype graph, find all components within the discordant graph, which are isomorphic (‘equivalent’) to the prototype graph (see Materials and Methods). To illustrate this concept, an isomorphism between the graphs and in Figure 2D is indicated. Since the prototype corresponds to a clearly defined SV, the mapping contains all information that are required to reconstruct the event exactly. In our example (Figure 2C–D) we can conclude that the genomic region between position Y and position Z was duplicated, inverted and upstream inserted at position X. Based on this framework, we defined a comprehensive set of prototype graphs (see Materials and Methods) and tested the performance of our algorithm on simulated data on 16 different SV types (see Text S1 and Table S4). Furthermore, we compared our algorithm to the recently published SV detection algorithms DELLY [33], BreakDancer [16], Pindel [14] and Clever [34]. The results of the simulation show that DSVD can identify all simulated SV types with high recall ranging from 0.88 (intrachromosomal translocation) to 0.96 (small deletions) for a coverage of 20. Although DSVD has a broader detection spectrum, the algorithm performs comparably to the other tested tools on the subset of SV types within their detection range (Figure 3). These simulation results indicate that the algorithm can be employed for the detection of SVs using sequencing data of the tumor and the control.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Performance comparison on simulated SVs.

Summary of the recall [%] achieved by DSVD, DELLY, BreakDancer, CLEVER and Pindel on different SV types (Text S1). The coverages specified during the read simulations are indicated in the legend. intrachr  =  intra chromosomal; interchr  =  inter chromosomal; down  =  downstream; up  =  upstream; no inv  =  no inversion; inv  =  inversion.

https://doi.org/10.1371/journal.pone.0087090.g003

Using concordant and discordant coverages to infer SV zygosity

The paired-end reads were subjected to the DSVD workflow outlined above. The maximum score values, obtained during the reconstruction of the fragment sequences, follow a bimodal distribution (Figure S4A). The mode around zero corresponds to non-overlapping read pairs derived from long fragments and the mode at one originates from overlapping read pairs. For subsequent analysis we only considered read pairs having a score equal to one, corresponding to a perfect overlap of at least 13 bases (see Materials and Methods). The reads pairs were subsequently aligned to the reference genome using a seed-based alignment (Figure 2A). The number of concordant and discordant alignments, obtained by the corresponding classification, is shown in Table 1. Then, we computed the overall read coverage – the genomic regions covered by concordant or discordant read pairs – to assess the fraction of the genome accessible to SV detection based on our sequencing data.

The overall read coverage encompasses 91% to 96% of the different euchromatic regions (Figure S4B). A comparison between read coverage and unique alignability (93% to 96%) indicates that the largest proportion of the non-repetitive part of the genome is covered (Figure S4B). Due to their repetitive nature, heterochromatic regions showed a substantially lower overall coverage (21% to 53%, Figure S4C). In order to visualize the SVs called by DSVD in the genome browser, we computed the concordant and discordant coverages, from the corresponding alignments of the tumor and the control (Table 1). The concordant and discordant coverage exhibit opposite characteristics (Figure 4A), i.e. a decrease in the concordant coverage generally coincides with an increase in discordant coverage, which is indicative for underlying structural differences to the reference genome. Furthermore, consideration of concordant and discordant coverages allow to infer the zygosity of an SV: while homozygous SVs result in loss of both reference wildtype alleles and, hence, no concordant read-pairs are expected (Figure 4B, I and II), in heterozygous events one wildtype allele remains to contribute to the concordant coverage (Figure 4B, III).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Concordant and discordant read coverage reveal extensive similarity between the tumor and control.

(A) A genome browser view of representative kb of the Drosophila reference genome. The tracks, denoted concordant and discordant, represent the total number of concordantly and discordantly aligning read pairs at a particular genomic position. (B) A browser view of homozygous and heterozygous SVs at higher resolution. In the case of homozygous SVs (I/II), the concordant coverage is decreasing to zero, as no wildtype allele is present anymore. In contrast, heterozygous events contain both a wildtype allele and an acquired SV, and are therefore characterized by a decrease within the concordant coverage to 50% (III). (C) Venn diagram representing the number of small insertions (of size bp) identified within the tumor and the control with weight (upper) and weight (lower), respectively. (D) Smoothed scatterplot representing the concordant coverage on the vertical axis and the discordant coverage (weight) on the horizontal axis for different subsets of small insertions. The three columns, from left to right, correspond to all small insertions (All), small insertions found in both genomes (Intersection) and small insertions specifically identified within the indicated genome (Specific), respectively.

https://doi.org/10.1371/journal.pone.0087090.g004

Genome instability is not a pre-requisite for neoplastic epithelial growth

Based on all discordant read pairs the discordant graph was constructed for the individual samples (Figure 2D). This led to identification of 311471 and 223285 distinct vertices (breakpoints) for the control and the tumor, respectively. To identify specific SVs within the discordant graphs we employed all previously defined prototype graphs in order to search the discordant graph for isomorphic subgraphs. Our search for small insertions, i.e. insertions which can be entirely characterized by a read pair (size ), revealed 73281 and 65890 events in the control and the tumor, respectively, of which 51383 are detectable in both genomes allowing a tolerance of one base to account for equivalent alignments (Figure 4C, see Materials and Methods). Moreover, this overlap further increases when the minimum required weight (, the number of read pairs supporting the same event) is raised from 1 to 4 (Figure 4C and Figure S5A). This tendency suggests that high confidence events are more likely to be present in both samples and, thus, to originate from germline rather than somatic mutations. The size distribution of insertions (Figure S5B) shows that most insertions are of length one, which is in good agreement with findings from rat, mouse and human, pointing to a median size of detected indels within introns of approximately 3 [35]. To assess the zygosity of small insertions, the number of concordant read pairs spanning the insertion site was considered (Figure 4D). This analysis identified three distinct classes (Figure 4D column 1–2) of small insertions characterized by: (i) high concordant and low discordant coverage (ii) comparable concordant and discordant coverage (iii) low concordant coverage and high discordant coverage. Insertions within the first class are likely to correspond to technical errors as they are characterized by low discordant coverage. In contrast, insertions belonging to the second and the third class are likely to correspond to heterozygous and homozygous events, respectively. Notably, the vast majority of small insertions specifically detected in either genome can be attributed to the first class (Figure 4D, last column) and therefore corresponds to low confidence events.

Next, we sought to identify putative mechanisms responsible for the formation of insertions. To this extent, we analyzed inserted sequences and their context of insertion. We found that the 10 most prevalent insertions are similar in frequency and sequence identity within the tumor and the control (Figure 5A), and that most of these insertions are A/T rich. Subsequently, for each single base insertion, we computed the number of insertions within simple repeats of the same base (Figure S5C), and compared it to the total number of insertions. Our analysis revealed that the largest fraction of single base A or T insertions localizes within simple A or T repeats, respectively (Figure 5B). This indicates that context specific DNA replication errors, such as replication slippage [36], may be causative since this mechanism is known to cause indels within simple repeat sequences [37]. In addition, replication slippage is known to induce errors during PCR amplification [37]. It is therefore possible that aberrant DNA fragments are generated at low frequencies during PCR amplification of the genomic DNA library, possibly explaining the sample-specific low-confidence SVs (Figure 4D and Figure S5A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Genomic context analysis can indicate mutational mechanisms causing SVs.

(A) DNA sequence and insertion frequency of the 10 most commonly inserted sequences identified within the control genome. For the tumor the tenth most frequently inserted sequence corresponds to CA with 897 insertions. For the sake of a clear representation the eleventh most frequently inserted sequence (AAA, 894 insertions) is shown. (B) The fraction of single base insertions within simple repeats consisting of the same base type, computed with respect to all single base insertions. Simple repeats of a minimum length of 4 were considered. (C) A genome browser view of a genomic locus containing two insertions (I/V), two deletions (II/IV) and one tandem duplication (III). As indicated by the discordant coverage and horizontal bars, these high-confidence SVs are both identified within the tumor and the control genomes, and have therefore been inherited from the parental strains.

https://doi.org/10.1371/journal.pone.0087090.g005

Another mechanism resulting in permanent structural changes at the target site is the conservative transposition of DNA transposons. Following the excision of the transposon, a direct repeat remains at the target site [38]. To investigate the potential relevance of this mechanism, we identified all small insertions longer than two bp exhibiting perfect sequence identity to either flanking sequence, a requirement for a direct repeat (Figure S6A). The analysis revealed 15073 and 12060 such events in the control and the tumor, respectively, of which more than 50% was identified within low-complexity A/T-rich simple repeats (Figure S6B), complicating the distinction between replication slippage and transposition. However, by analyzing non-simple insertions (more than 2 different bases) we identified 5795 and 5375 insertions in the control and the tumor (4765 in common), respectively, which correspond to potential transposon-mediated events.

Using the deletion prototype graph, we identified 71969 deletions in the control and 64737 in the tumor, which showed similar characteristics to the insertions. In total 41696 deletions were detected within both samples (allowing one base tolerance) and high confidence events are more likely to be common to the two genomes (Figure S7A–B). The size distribution of deleted sequences indicates that single base deletions occur at highest frequency (Figure S5B). As compared to insertions, deleted sequences show a similar A/T enrichment and also occur preferentially within simple repeats (Figure S7C–D), suggesting a similar mechanism of formation. Examples of deletions found in the tumor and the control are shown in Figure 5C (I/III) and Figure S8A–B.

To assess the effect of small insertions and deletions on coding sequences of cancer-related genes, we analyzed indels (weight ) which were specifically identified in the tumor or the control using VariantAnnotation (Obenchain et al., VariantAnnotation: Annotation of Genetic Variants, package version 1.4.5). This analysis revealed approximately 2.5 times more genes potentially affected in the control than in the tumor (Table S1 and S2). In addition, GO term analysis of the affected genes did not result in any tumor specific term enrichments (-value ) [39].

Searching for more complex SVs, revealed that tandem duplications and inversions occur at lower frequency and, similar to indels, the largest fraction of these events was common to both genomes (Figure S9). An exemplary tandem duplication is shown in Figure 5C (II). In addition, 15 and 13 translocations and insertional duplications were detected in the control and the tumor at a MRW of 2, respectively, of which 10 were common to both genomes.

Finally, visual inspection suggested that coding sequences are less susceptible to accumulate SVs, as they are characterized by a decrease in discordant coverage in comparison to non-coding regions (Figure 6A). To assess the breakpoint distribution globally, we partitioned the genome into distinct functional subsets and computed the normalized number of breakpoints therein. We found that exons, in particular the coding sequences, are generally less affected by SVs, whereas non-coding regions, such as intergenic or intronic sequences, exhibit a much higher predisposition for the accumulation of SVs (Figure 6B). These results are in accordance with previous studies [40]–[43] demonstrating the accumulation of SNPs and indels in intronic sequences of Drosophila.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 6. Coding sequences are less susceptible to SV accumulation.

(A) Genome browser view depicting the concordant and discordant coverage of the control (blue) and the tumor (red) samples across two protein-coding genes, and identified SVs therein. The detected insertions and deletions localize outside of coding sequences, and affect introns, intergenic spaces and UTRs. (B) Genome-wide breakpoint distribution across distinct functional compartments. Different subsets of the genome were selected according to following characteristics: genome corresponds to the full-length genome; the unique genes do not share common positions with any other gene; overlapping genes are non-unique genes; exonic regions, containing 3′UTRs, 5′UTRs and coding sequences (CDS) were obtained from the unique genes in order to avoid ambiguity; In addition, intronic and intergenic regions as well as donor/acceptor splice sites (splice sites) were considered. For each subset the number of contained breakpoints was computed and normalized to the total length.

https://doi.org/10.1371/journal.pone.0087090.g006

Experimental validation confirms SVs identified by DSVD

To experimentally validate our findings, we randomly selected 24 SVs of different type and size ranges, in addition to the two deletions (I/III) and the tandem duplication (II) in Figure 5C, for PCR-based analysis (Table 2 and Table S3). PCR was performed using genomic DNA of distinct individuals of the same genotype as the tumor tissue. Genomic DNA from the Drosophila reference genome strain was used as a control. In total, 23 reactions were confirmative as SVs were specifically detected within the tumor genome (Figure S10). However, the zygosity of four SVs differed from the predicted zygosity. Therefore, we extracted genomic DNA from ten individuals of the parental strains in order to retest these four events and, in addition, to retrace selected heterozygous SVs (Figure S11). Notably, and apart from one size deviation, all tested alleles were detected within parental individuals suggesting that the observed deviation from the predicted zygosity of SVs is most likely due to the non-isogenic background of the parental strains (Table 2) and Figure S11). This genetic heterogeneity can also explain that 4 out of 27 tested SVs were exclusively detected within the parental DNA, for which more individuals were tested.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of PCR-based validation experiments.

https://doi.org/10.1371/journal.pone.0087090.t002

Furthermore, to validate the predicted breakpoint locations, we performed Sanger sequencing with selected homozygous SVs (Figure S10). Sequencing confirmed that DSVD was indeed able to identify SV breakpoints with single bp resolution (Table S5).

Induction of tumors and image recording

Tumors and mutant cell clones were induced, transplanted and documented essentially as described in [26]. The utilized fly strains are available from the Bloomington Drosophila Stock Center at Indiana University, USA, and the Vienna Drosophila RNAi Center, Austria. The parental ph RNAi stock with the VDRC transformant ID10679 and the genotype [w1118; +; 10679/TM3, Sb] was outcrossed to a strain of the genotype [w1118; +; TM3, Sb/TM6, Tb] to establish [w1118; +; 10679/TM6, Tb]. Males of this genotype were then mated to virgin females of the previously established fly strain [P{w+mC = UAS-Dcr-2.D}1, w1118; P{w+mW.hs = en2.4-GAL4}e16E, P{w+mC = UAS-myr-mRFP}1, P{w+m* = NRE-EGFP.S}5A; +] to obtain neoplastic tumor (at 25°C) or control tissue (at 18°C) samples for sequencing from the offspring larvae. To record images, wing imaginal discs were attached to the surface of glass cover slips. The images depict RFP in red, EGFP in green, and bright field images in blue. For RFP and EGFP maximum projections and for the bright field images average projections of horizontal optical sections are displayed with identical magnifications (additional details in [26]). Brightness and contrast were adjusted using Photoshop. To increase visibility of the discs shape in Figure 1, the discs outline in the blue channel (bright field images) were selected in Photoshop, the selections inverted and the backgrounds replaced with black.

Preparation and sequencing of the genomic DNA library

Standard procedures were utilized throughout. Genomic DNA was isolated with the QIAamp DNA Micro Kit (Qiagen) and fragmented using the Bioruptor Plus (Diagenode); the Agilent High Sensitivity DNA Kit was used for quality control; libraries were prepared with the TruSeq DNA Sample Prep Kit (Illumina); to obtain a read length of 150 bases the TruSeq SBS Kit (v3) was used on a Illumina Rua HiSeq2000 equipped with a HiSeq Flow Cell (v3) and running the Illumina Pipeline Version 1.13.48.

Reconstruction of the fragment sequence

In order to determine the position of the overlap between mates, the forward strand of read 1 and the reverse complement of read 2 were considered to compute a position-dependent overlap score as:where

in which represents the base j (0-based offset) in read 1, represents base in the reverse complement of read 2, denotes the length the reads (here ) and represents the minimum required overlap of both reads. The starting index of the overlap can be computed as:

The probability of a false positive overlap is therefore equal to . In this study a conservative value of 13 was chosen.

Seed-based and global alignments

Seeds of the paired-end reads were aligned to D. melanogaster reference genome (dm3 BDGP Release 5) using Bowtie2 [47] with parameters -3 120 -N 1 -L 30 -k 11. Multiple alignments were discarded. The global alignment of the reconstructed fragment sequence was performed using the Needleman-Wunsch algorithm [48] with affine gap penalties [49] by setting gap opening penalty  = 10 and gap extension penalty  = 0.5.

Representation of structural variants using graphs

A graph is defined as a pair . The elements of the vertex set corresponds to distinct breakpoints. The edges in with represent directed connections between the breakpoints established by a pair oriented from read 1 to read 2. denotes the set of distinct edge types. An edge of type from node to is denoted as and the weight represents the number of read pairs supporting the same edge (not explicitly indicated for the sake of clarity in notation). The edge types belonging to are shown in Table 3. Edge type only occurs in combination with any other edge type. Since fragment orientation during sequencing is arbitrary, different edges may represent the same information. To account for such cases the following equivalence relations are defined on :

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Different directed edges belonging to . chr  =  chromosome.

https://doi.org/10.1371/journal.pone.0087090.t003

The existence of equivalent edges implies the existence of equivalent graphs. Hence, the equivalence class of a graph, i.e. the set of graphs corresponding to the same SV, was defined as follows. Consider the set of graphs , where and denote the vertex set and the edge set of , respectively. Two graphs are equivalent if there exists a bijection with and , where is defined by the equivalence relations (1–3) on the edges as:

Multiple optimal alignments of the reconstructed fragment sequence and the minimal reference frequently occur within repeat sequences and hamper exact breakpoint identification. To illustrate this problem, consider the insertion of a single A into a homopolymer stretch of As such as AAA. Here, the following cases are indistinguishable: AAAA, AAAA, AAAA and AAAA and would result in the identification of different breakpoints. The problem of multiple optimal alignments becomes more relevant during the construction of higher order graphs. Indeed, identification of complex SVs require simultaneous identification of the same breakpoints from independent global alignments of different fragment sequences. Failure in doing so can lead to the identification of distinct breakpoints impeding the construction of the complete graph. To overcome this problem and ensure identification of the same breakpoint by different fragments, DSVD exhaustively constructs a graph corresponding to each optimal alignment (including a tolerance of bases) resulting in collections of graphs representing equivalent SVs. From each collection one graph is selected. The selection is based on graph order and genomic position of the corresponding SV: higher order is prioritized over lower order and, in addition, graphs corresponding to central breakpoints within ranges of equivalent breakpoints are favored.

Identification of structural variations using prototypes

Graph prototypes corresponding to well defined SVs were defined for: small insertions, deletions, tandem duplications, inversions, insertional duplications and translocations. For the last two SV types the following cases were exhaustively considered: interchromosomal vs. intrachromosomal, downstream vs. upstream and with and without inversion of the inserted sequence. The prototypes were then used to find all components of the discordant graph which are isomorphic to any graph within the equivalence class of the prototype. An isomorphism from a graph to a graph is a bijection defined on the vertex sets as such that if and only if . The mapping is bijective and structure preserving as edges of the same type connect corresponding vertices within the inverse image and the image.

Implementation of the DSVD algorithm

The DSVD algorithm, detailed instructions as well as relevant documentation is available at:

http://www.bsse.ethz.ch/egg/software/index (11 June 2013, date last accessed). For all graph-related operations, DSVD depends on the python module NetworkX [50].

Accession number

The sequence data of this study have been deposited at the short read archive (http://www.ncbi.nlm.nih.gov/sra) under accession no. SRP017639.