TAREAN: a computational tool for identification and characterization of satellite DNA from unassembled short reads

Abstract

Satellite DNA is one of the major classes of repetitive DNA, characterized by tandemly arranged repeat copies that form contiguous arrays up to megabases in length. This type of genomic organization makes satellite DNA difficult to assemble, which hampers characterization of satellite sequences by computational analysis of genomic contigs. Here, we present tandem repeat analyzer (TAREAN), a novel computational pipeline that circumvents this problem by detecting satellite repeats directly from unassembled short reads. The pipeline first employs graph-based sequence clustering to identify groups of reads that represent repetitive elements. Putative satellite repeats are subsequently detected by the presence of circular structures in their cluster graphs. Consensus sequences of repeat monomers are then reconstructed from the most frequent k-mers obtained by decomposing read sequences from corresponding clusters. The pipeline performance was successfully validated by analyzing low-pass genome sequencing data from five plant species where satellite DNA was previously experimentally characterized. Moreover, novel satellite repeats were predicted for the genome of Vicia faba and three of these repeats were verified by detecting their sequences on metaphase chromosomes using fluorescence in situ hybridization.

INTRODUCTION

Satellite DNA (satDNA) is a class of repetitive DNA that is characterized by its genomic organization into long arrays of tandemly arranged units called monomers. The monomer sequences are typically hundreds of nucleotides long and highly homogenized (1). Although monomer length is often used to classify genomic tandem repeats as microsatellites (2–7 bp), minisatellites (tens of bp) or satellites (hundreds of bp), it appears that satellites are best distinguished by forming longer arrays (tens of kilobases up to megabases) concentrated in relatively few genomic loci, while micro- and mini-satellite arrays are much shorter and scattered across the genome. These differences in genomic organization probably reflect different amplification and homogenization mechanisms acting on these repeats (2–5). In the majority of eukaryotic genomes studied to date, satDNA was predominantly located in subtelomeric and centromeric chromosome regions, and the role of satDNA in centromere determination and function is the subject of ongoing research (6). In some organisms, such as higher plants, satellite repeats are also located in interstitial chromosome regions, forming prominent heterochromatic bands (7,8). The overall patterns of satDNA distribution revealed by fluorescence in situ hybridization (FISH) are frequently used in karyotype studies because they can provide markers for distinguishing morphologically similar chromosomes (9,10).

Investigation of satDNA or its utilization as a cytogenetic marker requires a priori knowledge of the nucleotide sequences of satellite repeats in the species of interest. However, satDNA is among the most dynamic components of eukaryotic genomes and its high evolutionary rate results in considerable sequence diversification, therefore most satellite repeat families are species- or genus-specific (1). Consequently, identification of satDNA by its similarity to known repeats from phylogenetically distant taxa is not possible. For these reasons, there has been continuous demand for efficient ab initio methods for satDNA identification. Satellite DNA acquired its name from density gradient centrifugation experiments, where it was discovered as a constituent of satellite bands formed due to its different buoyant density compared to the bulk of genomic DNA (11). Thus, density centrifugation was the first method of satDNA isolation, followed by other experimental approaches based, for example, on the presence of specific restriction sites in monomer sequences (12) or on the self-priming of tandemly repeated sequences in a modified PCR protocol (8). Although these methods led to identification of numerous satellite repeats, they are mostly limited to isolation of highly amplified repeats and biased towards those that can be distinguished by some property of their sequences, such as the presence of conserved restriction site. However, the satellites lacking these features may remain unnoticed.

An alternative to experimental methods for satDNA isolation is to identify their presence in genomic sequence data. Due to the introduction of next generation sequencing technologies, generating such data is no longer a limiting factor for genome investigation. Bioinformatics tools, such as Tandem Repeats Finder (TRF) (13), can then be used to search genomic sequences for tandem repeats including satellite DNA. As reviewed by Glunčić and Paar (14), TRF is a representative of string matching algorithms, which are utilized in a number of computational tools for tandem repeat prediction, along with alternative approaches based on nucleotide autocorrelation functions (15,16) and Fourier transforms (17). However, a common limitation of these tools is their need for long input sequences, spanning more than one repeat monomer. Although such long contigs are routinely available from whole genome assemblies, they often lack or are severely underrepresented for satellite repeats. This is because satellite repeats are extremely difficult to assemble due to their structure and high sequence homogeneity (18). Thus, the search for satellite repeats should ideally be performed in unassembled reads but this approach is hampered by relatively short length of the reads produced by most of the currently used NGS technologies.

The task of repeat identification from unassembled NGS reads has been addressed by the introduction of a similarity-based clustering algorithm which evaluates all-to-all sequence comparisons between whole genome shotgun reads (19,20). When applied to low-coverage (0.01–0.50×) genome sequencing data, there are almost no similarities detected between reads derived from single-copy sequences. On the other hand, reads that originated from repetitive elements produce multiple similarity hits and can thus be identified as clusters of frequently overlapping sequences. The number of reads in each cluster is proportional to the genomic abundance of the corresponding repeat, thus enabling its quantification. This repeat clustering analysis is at the core of the RepeatExplorer pipeline (20), which was originally designed and used for repeat characterization in plants (reviewed in (21)), but also proved to be efficient in repeat identification in other organisms, including bats (22), fish (23) and insects (24).

The clustering algorithm employed by RepeatExplorer represents the reads and their sequence similarities as nodes and connecting edges, respectively, in a virtual graph, and identifies read clusters by examination of the graph topology (19). In addition to efficient partitioning of the graph into clusters, this approach has the benefit of providing graphical representation of individual clusters. The shapes of these graphs reflect the genomic organization and sequence variability of corresponding repeats, ranging from linear structures typical for dispersed transposable elements to circular or globular shapes of tandemly repeated sequences (19). It has been demonstrated for a number of species analyzed using graph-based read clustering that the graph shapes can be reliably used to discover novel satellite repeats (25–31). However, the need to visually inspect the graph shapes represented a limitation of this approach and prevented its full automation. Another problem with this approach concerned identification of the most abundant variants of monomer sequences which are then needed for downstream applications, including a design of hybridization probes or PCR primers. Inferring monomer consensus using traditional methods based on multiple sequence alignments is not feasible due to large numbers of analyzed reads and a principally similar approach employing sequence assembly results in multiple contigs which require further manual processing. However, it was shown that alignment-free approaches utilizing k-mer frequency statistics are more suitable for monomer reconstruction from unassembled sequence reads (29,32,33) and therefore could fill this last gap in the automated workflow once implemented into an efficient computational tool.

In this work, we present tandem repeat analyzer (TAREAN), a computational pipeline which was built on the principles of graph-based repeat clustering, enhanced and supplemented with additional tools facilitating unsupervised identification and characterization of satellite repeats from unassembled sequence reads. The pipeline uses low-pass whole genome sequence reads as its input and performs their graph-based clustering as the first step in the analysis. Resulting clusters, representing all types of repeats, are then examined for the presence of circular structures characteristic for tandem repeats. This is achieved by constructing directed graphs from read similarities and selecting clusters that contain strongly connected components in their graphs. In addition, paired-end read information is utilized to discriminate clusters representing potential satellite repeats from other types of tandemly repeated sequences. Reads from these clusters are then decomposed to k-mers and fractions of the most frequent k-mers are used for reconstructing representative monomer sequences for each satellite repeat. To test the efficiency and specificity of the pipeline, we first analyzed NGS data from five plant species with various numbers of previously characterized satellite repeat families which differ in their genomic abundance. Moreover, we demonstrated that TAREAN can also identify novel satellite repeats that were subsequently verified by their detection on metaphase chromosomes using FISH with probes designed according to reconstructed monomer sequences.

MATERIALS AND METHODS

The workflow of TAREAN

Input data. The analysis requires paired-end reads generated by whole genome shotgun sequencing provided as a single FASTA formatted file. Read length should be 100–200 nt and the number of analyzed reads should represent less than 1× genome equivalent (genome coverage of 0.01–0.50× is recommended). Illumina 2 × 100 nt reads were used in this work, however, paired-end reads generated by other NGS platforms should be also suitable for analysis, provided that the sequenced fragments are of sufficient length to avoid frequent overlaps of paired-end read sequences. Reads should be of uniform length, quality-filtered (quality score ≥10 over 95% of bases, no Ns allowed) and only complete read pairs should be submitted for analysis. The analysis workflow is schematically depicted in Figure 1.

Figure 1.

Schematic representation of the TAREAN analysis workflow.

Graph-based clustering

The read clustering algorithm is the same as described by Novák et al. (19). Briefly, reads are subjected to all-to-all sequence comparisons and their mutual similarities exceeding a specified threshold (90% similarity over at least 55% of the read length) are represented as a graph in which vertices correspond to sequence reads and overlapping reads are connected by edges. The resulting graph is then subjected to the Louvain method for community detection and partitioned into clusters (34) (Figure 1A). Although this method is computationally efficient and does not generate chimeric clusters, its drawback is that some families of repetitive elements frequently get split into multiple clusters rather than being represented as a single cluster (19,30). However, by utilizing paired-end read information, these split clusters can be identified and merged. Merging is performed for clusters that share significant proportions of broken read pairs (that is, when paired-end reads are present in different clusters), as determined with the formula:

where W is the number of read pairs shared between clusters x and y, and n_xand n_y are the numbers of broken read pairs in clusters x and y, respectively. The cutoff for cluster merging used in our analysis was set to k_x,y ≥ 0.2.

Automated detection of circular structures in cluster graphs

Following clustering, each cluster that represents abundant genomic repeat is examined for the presence of circular structures indicative of tandem repeats (clusters that contain at least 0.01% of input reads are analyzed by default). This is achieved by constructing a directed graph from the read similarities (Figure 1B) and testing if the graph is strongly connected, which means that it can be traversed from one read to any other read through a series of similarity overlaps (35). This is implemented by first constructing an edge signed graph Σ where vertices represent reads, edges connect overlapping reads and signs of the edges reflect orientation of the overlapping reads, being positive for forward to forward and negative for forward to reverse complement overlaps (36). The minimum spanning tree Σ_msp of the graph Σ is then traversed using depth first search and each vertex which is connected to a previously visited vertex with a negative edge is switched (i.e. reverse-complemented). The resulting switching equivalent graph of Σ_msp is used in the next iteration, finally leading to the directed graph G where all edges are positively signed. Next, the proportion of the largest strongly connected component in graph G is calculated as the connected component index C:

where V(G) is the number of vertices of the graph G and V(G_LSCC) is the number of vertices in the graph G_LSCC, which is a subgraph of G and corresponds to its largest strongly connected component (Figure 1C). For graphs derived from exact tandem repeats, C = 1.

Identification of putative satellite repeats

Although the parameter C facilitates the identification of clusters representing tandemly repeated genomic sequences, it does not efficiently discriminate clusters derived from satellite DNA from those representing other types of tandem repeats. Therefore, an additional cluster characteristic providing a proportion of broken read pairs is calculated. A typical feature of satellite repeats is that they occur in long contiguous arrays of monomers ranging up to megabases in length, whereas other tandem repeats form arrays in a range of hundreds to thousands of bp. Consequently, clusters of satDNA contain low proportions of broken read pairs, because most sequenced DNA fragments are entirely made of the same repeat. On the other hand, the proportions of broken pairs are much higher in tandem repeats scattered in the genome in a high number of short arrays, because many sequenced fragments span the junctions between a tandem repeat array and its neighboring genomic sequences. This is evaluated as the pair completeness index P using a formula:

where N_C is the number of complete read pairs in the cluster N and N_I is the number of broken pairs. Both criteria, C and P, are then used simultaneously to detect putative satellite repeats, which have expected values close to 1 for both. Estimation of the threshold values of C and P suitable for sensitive yet reliable identification of putative satellite repeats was performed by re-analyzing 2968 manually annotated clusters from 11 plant species selected from the dataset published by Macas et al. (30). The estimation was done using discriminant analysis based on a Gaussian finite mixture model (37) as implemented in the R package mclust.

Reconstruction of monomer sequences from the most frequent k-mers

Reconstruction of prevailing sequence variants is performed by counting the occurrences of k-mers in a set of oriented reads obtained from the directed graph G. k-mers with lengths k = 11–27 are analyzed in parallel. The use of oriented reads ensures that the sequence reconstruction will be performed in one direction only, avoiding parallel reconstruction of its reverse complement. Identified k-mers are sorted based on their proportions in the analyzed sequence data and the resulting sorted list with the most frequent k-mers at the top is used in the subsequent analysis. The most frequent k-mers that represent 50% of the sequence data are used to construct a de Bruijn graph B and are removed from the k-mer list. k-mer frequencies are represented in the graph as weights of the corresponding vertices. The graph is then checked for the presence of cycles, and the subgraph B_LSCC with the largest strongly connected component is identified. If there is no strongly connected component or if the sum of vertex weights in B_LSCC is less than the threshold p_km, additional k-mers from the top of the list are iteratively added until the threshold is reached (the optimal value of p_km was tested empirically and set to 0.225). This process leads to graphs with reduced numbers of vertices yet containing cycles corresponding to prevalent monomers of tandem repeats (Figure 1D). Variants of monomer sequences are then extracted from the cycles by converting the sequences of k-mers making up de Bruijn graphs to nucleotide sequences, aligning sequences of the same length and calculating consensus and position probability matrices (PPM) from k-mer weights (Figure 1E). To limit the number of cycles used for monomer reconstruction, only the cycle with the highest weight (the sum of weights of all vertices in the cycle) is considered for each graph branch. In case of length variation in the reconstructed monomer sequences, multiple PPMs are produced and the one with the highest total weight is reported.

Identification of other types of repetitive sequences

Genes coding for 45S and 5S ribosomal RNAs are arranged as multi-copy tandem arrays in eukaryotic genomes and as such are detected as putative satellites by TAREAN. However, they are identified by similarity searches to a custom database of Viridiplantae rDNA sequences and reported separately in the program output. A specific type of repeats with potential for producing false-positive results are LTR-retrotransposons, which, due to the presence of direct terminal repeats, form circular graphs (19). To avoid this misclassification, we search for LTR retrotransposon-specific features in the reconstructed consensus sequences, including the presence of primer binding sites (PBS) complementary to some tRNAs (38) and retrotransposon protein-coding open reading frames longer than 300 bp.

Implementation

TAREAN is implemented using custom python and R scripts. Graph analysis was performed using igraph, a software collection for complex network research (39). All scripts and databases are available for download from http://w3lamc.umbr.cas.cz/lamc/resources.php. Additionally, TAREAN was implemented under Galaxy web-based environment (40) and made available as a tool in the public RepeatExplorer server (20) at http://www.repeatexplorer.org. The analyses presented in this paper were performed on Linux-based servers equipped with 16 GB RAM and 4–16 CPUs.

Pipeline testing and validation of the results

Data

The pipeline was tested using genomic shotgun Illumina reads from five species with previously characterized satellite repeats. The reads were downloaded from European Nucleotide Archive (http://www.ebi.ac.uk/ena) under accession numbers ERX379412 (Vicia faba L.), ERR063464 (Pisum sativum L.), ERP001569 (Luzula elegans Lowe), PRJEB9643 (Rhynchospora pubera (Vahl) Boeckeler) and SRX118541 (Zea mays L.).

Experimental validation of predicted satellite repeats

Reconstructed consensus sequences of satellites predicted by TAREAN in Vicia faba were used to design oligonucleotide probes for fluorescence in situ hybridization: Vf_TA11_H2, biotin-5΄-GGT TAC TTC ATC ACT AAG AAA CTA AGT TAA AAG ACT ATT AMT TAA TGA CAC-3΄; FokI_H1, fluorescein-5΄-CTA CCT TCC ATA ATG ACA AGG CTA CCA TCC ATT GGA GTA ACA AAA ATC TC-3΄. The oligo-probes were labeled with biotin or fluorescein at their 5΄ ends during synthesis. Alternatively, PCR primers were designed for amplification and cloning of satellites with longer monomers: Vf_TA39_1, 5΄-AGC ACG AAT AAA ACT AAA GTT C-3΄; Vf_TA39_2, 5΄-TAC TTT TGA AGT GAA ATG GAG-3΄; Vf_TA157_1, 5΄-GGT ATG AGA ATG GTG TAT CTT TTA TCA-3΄; Vf_TA157_2, 5΄-AGA AAA GAT ATT TGG TTT CGA ATG A-3΄. All oligonucleotides were synthesized by Integrated DNA Technologies (Leuven, Belgium). Probe amplification from total genomic DNA of V. faba and cloning was performed as described in Macas et al. (30). Probes were labeled with biotin-16-dUTP (Roche Diagnostics GmbH, Mannheim, Germany) or Alexa Fluor 568 (Thermo Fisher Scientific, Waltham, MA, USA) using nick translation (41) and FISH was performed according to Macas et al. (42). The oligo-probe FokI_H1 specific for FokI satellite (43) was used for simultaneous hybridization (two-color FISH) with the novel repeats to provide characteristic banding patterns allowing the discrimination of all chromosomes within the V. faba karyotype (44). Chromosomes were counterstained with DAPI and examined using a Nikon Eclipse 600 microscope. Images were captured using a DS-Qi1Mc cooled camera and NIS Elements 3.0 software (Laboratory Imaging, Praha, Czech Republic).

RESULTS

Major features of the pipeline and estimation of optimal parameters

The TAREAN pipeline takes paired-end NGS reads as input and outputs a list of clusters identified as putative satellite repeats, their genomic abundance and various cluster characteristics. The lengths and nucleotide sequences of reconstructed monomers are also provided and are accompanied by a detailed output from k-mer-based reconstruction including sequences and sequence logos of alternative variants of monomer sequences. A summary of this information is provided in HTML format and includes a table listing all analyzed clusters (an example of the HTML output is provided as Supplementary Data). More detailed information about clusters is provided in additional files. When the analysis is performed on a Galaxy server, all generated results are downloadable as a zip archive. Since read clustering results in thousands of clusters, the search for satellite repeats is limited to a subset of the largest clusters corresponding to the most abundant genomic repeats. The pipeline is set to analyze all clusters representing at least 0.01% of the input reads, but this size threshold can be changed in order to adjust the sensitivity of the analysis. Besides the satellite repeats, three other groups of clusters are reported in the output (i) LTR-retrotransposons, (ii) 45S and 5S rDNA and (iii) all remaining clusters passing the size threshold. As categories 1 and 2 contain sequences with circular graphs, their consensus is calculated in the same way as for the satellite repeats.

Since two cluster characteristics, the connected component index C and the pair completeness index P, are crucial for identification of satellite repeats, we searched for their optimal cutoff values by evaluating a pool of 2968 clusters from 11 species of legume plants. These clusters were manually annotated during our previous study (30) and included 174 satellites; the remaining clusters represented other kinds of genomic repeats. The C and P values of these clusters were used as training data for discriminant analysis to find the best model for satellite prediction (Figure 2A). Clusters identified as satellites according to this model were denoted as high-confidence satellites. Additionally, we also chose less strict criteria of P > 0.4 and C > 0.7 to be able to detect less typical satellite sequences which are then reported as low-confidence satellites. Examples of clusters with different P and C values with corresponding graph shapes are shown on Figure 2B E. In the model-based prediction using discriminant analysis, 143 (82%) of the reference satellites clusters were correctly classified as high-confidence satellites with a false positive rate of 1.4% (Table 1). Employing the low-confidence category criteria resulted in detection of 173 out of 174 control satellite clusters but the higher sensitivity led to an increased false positive rate (18%).

Figure 2.

Training dataset and examples of cluster graphs. (A) Scatter plot of C (connected component index) and P (pair completeness index) values for all reference clusters. Red dots mark clusters that were manually annotated as satellite repeats. Threshold for classification, based on the best discriminant analysis model, is shown as a blue line and defines the high-confidence satellite group. Green lines mark empirically selected thresholds for the low-confidence category. (B–D) Examples of repeat clusters visualized as graphs where nodes represent sequence reads and edges connect reads with sequence similarities. Nodes belonging to the largest strongly connected components of the graphs are red; corresponding C and P values are shown below each graph.

Table 1. Performance of automatic classification model. A confusion matrix of numbers of clusters annotated as satellite repeats compared to the reference obtained by manual annotation.

		Manual annotation (reference)
		Non-satellite	Satellite
Automatic (model)	Non-satellite	2756	31
	Satellite	38	143

Testing the pipeline performance using previously characterized satellite repeats

To validate the pipeline sensitivity and accuracy, we analyzed NGS data from five plant species in which satellite DNA was previously experimentally characterized (Table 2). The satellites were identified in these species using restriction digestion-based cloning and/or library screening (Z. mays, V. faba; (43,45–48)) or they were identified using bioinformatics tools, but subsequently verified by cloning, sequencing and FISH analysis (Rhynchospora pubera, Pisum sativum, Luzula elegans; (25–27, 42)). These control species were selected for carrying diverse satellites with different monomer lengths, sequence variability, abundance and location in the genome. Moreover, these species represented three different types of chromosome organization, including species with monocentric (Z. mays, V. faba), meta-polycentric (P. sativum) and holocentric chromosomes (L. elegans, R. pubera).

Table 2. Evaluation of TAREAN performance in species with previously characterized satellites. Successful detection is marked by ‘++’ (high-confidence) or ‘+’ (low-confidence).

Species			TAREAN 500k			TAREAN max
Satellite				Merge			Merge
	Monomer [bp]	Abundance [% genome]	Coverage	NO	0.2	Coverage	NO	0.2	Monomer [bp]	Notes
Zea mays
Zea/Tripsacum	180	2.10	5775	–	++				180	(47)
CentC	156	0.20	635	++	++				156	(45)
TR-1(knobs)	350/180	0.11	156/303	++	++				359	(46)
Rhynchospora pubera										(26)
Tyba-1	171	1.80	5211	+	+				172
Tyba-2	171	1.16	3358	+	+				172
Luzula elegans										(27)
LeSAT4	190/220/360	2.40	6253/3300	+	++	25263/13333	+	++	170
LeSAT11	56	1.10	9723	++	++	39286	++	++	56
LeSAT7	75	0.97	6402	++	++	25867	++	++	75
LeSAT16	178/195	0.82	2280/2028	++	++	9213/8410	++	+	177/195
LeSAT18	Variable	0.55	n.a.	++	++	n.a.	++	++	56
LeSAT23	57	0.50	3397/1037	++	++	17544	++	++	57
LeSAT17	161	0.48	1476	–	–	5963	+	–	161
LeSAT38	137	0.37	1337	–	–	5401	–	–	–
LeSAT25	6	0.36	29700	–	–	120000	–	–	–	SSR
LeSAT22	51/167	0.35	1037	++	++	13725/4192	++	++	51
LeSAT28	390/730	0.32	406/217	++	++	1730/877	++	++	392
LeSAT43	190	0.23	599	++	++	2421	++	++	189
LeSAT9 + 21	43	0.22	2533	++	++	10233	++	++	43
LeSAT63	90	0.13	715	++	++	2889	++	++	89
LeSAT72	4	0.13	16088	–	–	65000	–	–	–	SSR
LeSAT36	6	0.12	9900	–	–	40000	+	+	24	SSR
LeSAT99	180	0.11	303	+	+	1222	++	++	180
LeSAT109	33	0.08	1245	++	++	5030	++	++	33
LeSAT89	41	0.06	724	–	–	2927	–	–	–
LeSAT27	42	0.06	672	++	++	2714	++	++	84
Pisum sativum										(25,42)
PisTR-B	50	1.37	13740	++	++	39516	++	++	50
TR-5	54	0.51	4731	++	++	13608	++	++	54
TR-2	440	0.21	235	–	–	677	–	–	–
TR-4	172	0.20	581	+	+	1672	+	+	173
TR-7	164	0.14	412	++	++	1184	++	++	164
TR-11	510	0.10	101	–	–	290	+	+	459	Low coverage
TR-3	81	0.06	358	++	++	1030	++	++	82
TR-19	2094	0.03	8	–	–	23	–	–	–	Low coverage
TR-1	867	0.02	12	–	–	35	+	+	866	Low coverage
TR-18	1644	0.01	4	–	–	11	–	–	–	Low coverage
TR-17	191	0.01	31	++	++	90	+	+	191	Low coverage
TR-6	245	0.01	22	–	–	65	–	–	–	Low coverage
TR-10	659	0.01	8	–	–	22	–	–	–	Low coverage
Vicia faba
FokI	59	3.20	26847	++	++	53695	++	++	59	(43)
pVf7	168	0.33	972	+	+	1945	+	+	169	(48)

The columns show (from left to right) previously reported monomer sizes and genome abundance of reference satellite repeats and results of their detection by TAREAN with 500 000 reads (‘TAREAN 500k’) and with maximal number of reads that could be analyzed (‘TAREAN max’). The last ‘Monomer’ column provides lengths of consensus monomer sequences reconstructed by TAREAN. Multiple values reflect several repeat variants differing in monomer length.

TAREAN runs were performed with 500 000 input reads which should provide sufficient sensitivity towards abundant satellites, yet keeping the computation time in the range of hours. Two analysis conditions were tested: the cluster merging option was either disabled or enabled at the cutoff value of 0.2. This analysis led to the successful detection of highly amplified satellites previously reported for Z. mays, V. faba and R. pubera (Table 2). In R. pubera, both previously characterized subfamilies of the same satellite (Tyba-1 and Tyba-2) sharing ∼70% similarity (26) were detected and distinguished. In the other two species, Luzula elegans and Pisum sativum, the analysis identified all highly abundant satellites with genome proportions exceeding 0.5%, but failed to detect some less-amplified satellites with estimated genomic proportions between 0.01 and 0.50% (Table 2). Thus, additional runs were performed with 2 million reads for L. elegans and 1.44 million for P. sativum, representing the maximal numbers of reads that could be processed at the given hardware configuration (the read numbers are different as they depend on the numbers of similarity hits between the reads, reflecting different proportions of repeats in each species). Although processing more reads improved the detection of four satellite repeats, 9 of 33 control satellites in these species remained unidentified. An investigation of the properties of these repeats provided an explanation of these results and enabled understanding of the sensitivity limits of TAREAN analysis, which were mostly determined by sequencing coverage, sequence homogeneity of monomers, genomic organization and similarities of satellites to other genomic repeats.

The sequencing coverage of a satellite was calculated as the total length of reads covering its sequences divided by monomer length. Thus, the calculation of coverage provided a normalization for genomic abundance (%) values, because satellites with the same genomic proportions but differing in monomer length have different coverages. For example, the P. sativum satellites, TR-17 and TR-18, both had genome proportions of 0.01%, but the former had higher coverage due to its shorter monomer length. The group of these less-amplified satellites (labeled as ‘low-coverage’ in Table 2) revealed that the sensitivity limit of the analysis was at the coverage range of 30–100×, probably depending on the sequence homogeneity of individual satellites. Thus, the failure to detect the P. sativum satellites TR-18, TR-6, TR-10 and TR-19 could be explained by their low abundance. Moreover, in the case of TR-19, the detection was also hampered by the fact that this repeat represents a longer monomer variant of TR-11 and thus the two repeats occurred in the same cluster. Since TR-19 has a lower genomic copy number compared to TR-11 (25) its monomer was not reconstructed and reported.

It has been reported that some satellite repeats originated by amplification of short tandem arrays present in other genomic repeats such as retrotransposons (49). Such satellites may be difficult to detect by TAREAN because their cluster graphs could contain substantial portions of non-circular components representing neighboring regions of repeats from which they originated. This was the case in the L. elegans satellite, LeSAT38, and in TR-2 in P. sativum. Another group of undetected satellites comprised three L. elegans repeats with extremely short monomers corresponding to simple sequence repeats of 4 bp (LeSAT72) or 6 bp (LeSAT25 and LeSAT36). These repeats failed to produce clusters due to active masking of low complexity regions during sequence similarity searches. The partial exception was LeSAT36 where the basic motif of 6 bp also occurred as a mutated higher order repeat of 24 bp which was reported by TAREAN (Table 2).

Identification and verification of novel satellite repeats

In addition to detecting previously described repeats, there were additional putative satellites reported for some of the analyzed species. The highest number was 12 novel satellites, which were identified for V. faba in the run analyzing a maximum number of 990 000 reads (Table 3 and Supplementary Data). One of them, Vf_TA70 was partially similar to the VicTR-B satellite described in several other Vicia species (8). No significant similarities were found for the other novel satellites.

Table 3. Putative novel satellite repeats identified in Vicia faba.

Satellite	Monomer [bp]	Genome proportion [%]	Copy number /1C	Notes
Vf_TA_11	191	1.20	843 000	Verified by FISH (Figure 3A)
Vf_TA_39	702	0.29	55 400	Verified by FISH (Figure 3B)
Vf_TA_62	687	0.15	29 300
Vf_TA_70	38	0.12	423 000	Similar to VicTR-B
Vf_TA_108	1482	0.05	4500
Vf_TA_109	870	0.05	7700
Vf_TA_123	878	0.03	4600
Vf_TA_137	603	0.03	6700
Vf_TA_143	352	0.02	7600
Vf_TA_154	560	0.02	4800
Vf_TA_157	781	0.02	3400	Verified by FISH (Figure 3C)
Vf_TA_158	313	0.02	8600

Three of the novel repeats, Vf_TA11, Vf_TA39 and Vf_TA157, differing in monomer length and genomic abundance (Table 3) were chosen for experimental validation by detecting their hybridization patterns on V. faba metaphase chromosomes using FISH. In all three repeats, band or dot-like patterns of signals typical for satellite DNA were detected. The repeat Vf_TA11 represented a highly amplified satellite with a 191 bp monomer, which was estimated to occur in 843 000 copies per haploid genome. The oligonucleotide probe (49 nt), designed according to the most conserved part of its reconstructed monomer sequence, produced a strong signal on the satellite arm of metacentric chromosome 1, and a number of minor signals close to centromeres of all acrocentric chromosomes (Figure 3A). Since monomer sequences predicted for the other two satellites were too long to be covered by oligonucleotide probes (701 and 782 bp), FISH was performed using PCR-amplified and cloned genomic fragments. In both cases the PCR with primers designed according to reconstructed monomer sequences yielded specific bands of expected lengths and their sequences confirmed predictions made by TAREAN (95.5% and 93.6% similarity to reconstructed consensus of Vf_TA39 and Vf_TA157, respectively). The resulting FISH patterns of Vf_TA39 consisted of multiple intercalary bands on most chromosomes, in agreement with relatively high abundance (55 400 copies/1C) of this repeat in the genome (Figure 3B). As expected, the much less amplified satellite Vf_TA157 (3400 copies/1C) produced weaker labeling which was limited to a single locus on chromosome 3 (Figure 3C).

Figure 3.

FISH localization of novel satellite repeats on metaphase chromosomes of Vicia faba. The probes for the novel satellites, Vf_TA11 (panel A), Vf_TA39 (panel B) and Vf_TA157 (panel C), are green, FokI repeats used for chromosome discrimination are labeled red and chromosomes counterstained with DAPI are blue.

DISCUSSION

In this work, we have introduced and validated TAREAN, a computational pipeline for the automated identification of satellite repeats from unassembled NGS reads. Although there are a number of computational tools available for detecting tandem repeats in assembled genomic sequences (50,51), corresponding tools utilizing short sequence reads are scarce. To our best knowledge, only two algorithms, DExTaR and MixTaR (52,53) have been published to address this problem. The former was designed for the detection of tandem repeats from de Bruijn graphs constructed for the purpose of genome assembly. It uses parts of de Bruijn graphs that were omitted from assembly and detects potential tandem repeats in the form of cycles. The method requires previous global assembly by a de Bruijn assembler such as ABySS (54) and is limited to the identification of exact tandem repeats. MixTaR represents an improved approach allowing detection of approximate tandem repeats, however, it requires long PacBio reads in addition to short Illumina reads for its analysis. Moreover, the algorithm was tested for detecting repeats with monomers up to 100 bp only, while most satellite DNA families have longer monomers (1).

In our previous work, we demonstrated that an alternative approach based on graph representations of repeat populations in eukaryotic genomes can be utilized for the identification of satellite repeats (19,29,31). This method, employing the RepeatExplorer pipeline (20) for performing similarity-based repeat clustering and generating graph visualizations, allows identification of approximate tandem repeats of any length, provided they are sufficiently represented in the analyzed short reads to form recognizable circular structures in their cluster graphs. Consequently, satellite repeats with various degrees of sequence conservation, and monomer lengths up to 5 kb can be identified (25,30,55). Recently, a modification of this approach, employing iterative clustering in order to improve its sensitivity towards low-copy tandem repeats, has been published by Ruiz-Ruano et al. (28). Nevertheless, both setups require human intervention for graph shape examination and the former does not provide consensus sequences of identified satellite repeats; features that were fully addressed in TAREAN.

Testing TAREAN performance using short NGS reads from five control species revealed its excellent efficiency in detecting highly abundant satellite repeats and very good performance in identifying less-amplified satellites. In addition to the repeat identification, consensus monomer sequences were accurately reconstructed in most cases. On the other hand, a fraction of previously described repeat families was not identified in test runs. When evaluating TAREAN performance, it should be acknowledged that the tool was specifically designed for the detection of genuine satellite repeats, a category of tandemly repeated sequences characterized by the forming of long contiguous arrays of highly homogenized monomer sequences. However, this category is not always clearly separated from other genomic tandem repeats. For example, some satellite repeats originate through the amplification of short tandem repeat arrays present in mobile elements. Thus, the same monomer sequences occur in the genome as short, dispersed tandem arrays, as well as in a few long arrays typical for satellite repeats (49). Such repeats then produce cluster graphs with intermediate features combining circular and linear structures, thus hampering their identification. Satellite repeats derived from a large intergenic spacer (IGS) of 45S rDNA represent a similar case, being present as short arrays within IGS and as amplified satellites elsewhere in the genome (48,56). The repeat pVf7 from Vicia faba (48) represented IGS-derived satellites in our data; although it was successfully identified by TAREAN, it was only listed in the low-confidence category (Table 2) due to its complex graph structure. Another example of a satellite with intermediate features that was reported with lower confidence was the Tyba satellite of Rhynchospora pubera (Table 2). Tyba is the satellite associated with centromeric chromatin, which is dispersed along holocentric R. pubera chromosomes. Thus, Tyba is organized in multiple arrays only up to tens of kilobases long as revealed by FISH and sequence analysis of BAC clones (26).

Regarding the ability to detect less abundant satellite repeats, there is no simple rule that could be used to determine a TAREAN sensitivity threshold. This is because the successful identification of a particular repeat depends on multiple factors, including its copy number in the genome, sequence variability, genomic organization and number of reads that were analyzed. In principle, increasing the number of analyzed reads results in more efficient detection of less amplified satellites (Table 2) but the genome sequencing coverage should not exceed 0.5-1.0x in order to avoid similarity hits between single-copy sequences during clustering analysis. However, in species with large and repeat-rich genomes, such coverage might be hard to reach due to constraints imposed by computational resources. The limiting factor for read clustering analysis is the number of similarity hits between reads, which increases with increasing proportions of high copy number repeats in the genome. Therefore, smaller numbers of reads can be clustered in highly repetitive genomes compared to those with low proportions of highly repeated sequences (19). On the other hand, even the low coverage used in this study for the large, repeat-rich genome of V. faba (990 000 reads correspond to 0.007× genome equivalent) proved to be sufficient to identify relatively rare repeats like Vf_TA157 with only thousands of copies per haploid genome (Table 3, Figure 3C).

AVAILABILITY

Command-line version of TAREAN can be downloaded from http://w3lamc.umbr.cas.cz/lamc/resources.php. The pipeline can also be run via Galaxy web interface at our public RepeatExplorer server (http://www.repeatexplorer.org).

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.

FUNDING

Czech Ministry of Education, Youths and Sports [LM2015047]; Czech Science Foundation [BP501/12/G090]; Czech Academy of Sciences [RVO:60077344]. Funding for open access charge: Czech Ministry of Education, Youths and Sports [LM2015047].

Conflict of interest statement. None declared.