Edinburgh Research Explorer Detection of copy number variants in African goats using whole genome sequence data

Background: Copy number variations (CNV) are a significant source of variation in the genome and are therefore essential to the understanding of genetic characterization. The aim of this study was to develop a fine-scaled copy number variation map for African goats. We used sequence data from multiple breeds and from multiple African countries. Results: A total of 253,553 CNV (244,876 deletions and 8677 duplications) were identified, corresponding to an overall average of 1393 CNV per animal. The mean CNV length was 3.3 kb, with a median of 1.3 kb. There was substantial differentiation between the populations for some CNV, suggestive of the effect of population-specific selective pressures. A total of 6231 global CNV regions (CNVR) were found across all animals, representing 59.2 Mb (2.4%) of the goat genome. About 1.6% of the CNVR were present in all 34 breeds and 28.7% were present in all 5 geographical areas across Africa, where animals had been sampled. The CNVR had genes that were highly enriched in important biological functions, molecular functions, and cellular components including retrograde endocannabinoid signaling, glutamatergic synapse and circadian entrainment. Conclusions: This study presents the first fine CNV map of African goat based on WGS data and adds to the growing body of knowledge on the genetic characterization of goats.


Background
Structural variations (SV) are an important source of genetic variation [1][2][3][4]. SV are generally considered to comprise a myriad of subclasses that consist of unbalanced copy number variants (CNV), which include deletions, duplications and insertions of genetic material, as well as balanced rearrangements, such as inversions and interchromosomal and intrachromosomal translocations [5]. Deletions and insertions are referred to as unbalanced SV because they result in changes in the length of the genome. Insertions or deletions in the genome are typically considered CNV when they are at least 50-1000 base-pairs (bp) long [6][7][8][9][10][11]. CNV are not as abundant as single nucleotide polymorphisms (SNP), but because of their larger sizes, they may have a dramatic effect on gene expression in individuals [12]. Duplication or deletion in or near a gene or the regulatory region of the gene may lead to modification of the function of the gene.
CNV cover about 4.5-9.8% of the human genome [13] and are associated with many Mendelian disorders [12]. Girirajan et al. [14] found that CNV significantly determine the severity and prognosis of many genetic disorders. Approximately 14% of diseases in children with intellectual disability are caused by CNV [15]. On the other hand, some CNV have been found to be associated with adaptive fitness of individuals, such as adaptation to starch diets associated in the gene encoding α-amylase [13].
Traditionally, microarray-based comparative genomic hybridization (array CGH) or SNP genotyping arrays are used to detect CNV. Several studies have been carried out using these methods to detect and map CNV in the goat genome, including studies by Fontanesi et al. [16] in four goat breeds; Nandolo et al. [17] in 13 East African goat breeds; and Liu et al. [18] in the global goat population.
Detecting CNV using array CGH and SNP genotyping arrays suffers from shortcomings that include hybridization noise, limited coverage of the genome, low resolution, and difficulty in detecting novel and rare mutations [19][20][21]. The development of whole-genome sequencing (WGS) technologies has made it possible for more rigorous and accurate detection of CNV.
According to Mills et al. [22], WGS-based CNV detection methods fall into four major approaches: methods based on paired-end (PE) mapping, split reads (SR), read depth (RD) and de novo assembly of a genome (AS). The PE and SR methods are useful for detection of small-scale CNV [23], and several algorithms are loosely based on them, including BreakDancer [24], Pindel [25], and Delly [26]. RD approaches are very useful for detection of larger CNV. Algorithms using this approach include CNV-Seq [27], CNVnator [28] and the event-wise testing approach (EWT) developed by Yoon et al. [29]. The methods can also be combined. For example, LUMPY [30] is able to combine two or more of the previous approaches to refine SV detection. Assembly-based approaches are computationally intensive and are therefore not generally used with WGS data [23,31]. Most of these SV-detection algorithms have been extensively reviewed [1,[31][32][33][34].
LUMPY implements a breakpoint prediction framework, where a breakpoint is defined as a pair of genomic regions that are adjacent in a sample, but not in the reference genome. The location of the breakpoint is determined using a probability function that considers different sources of evidence supporting the existence of a breakpoint, including information from discordant read pairs and split reads. A discordant read pair occurs when sequence from two ends of an insert are inconsistent when compared to the reference genome. These inconsistencies result from differences between mapping distance or the orientation between the pairs of sequences [35,36]. Split reads are sequences that map to the reference genome on one end only, and, as explained by Ye and Hall [33], such reads can indicate the location of a breakpoint with a high degree of certainty. There are similar algorithms that rely heavily on the use of breakpoints to determine genome rearrangements at single-nucleotide resolution, including Delly [26] and Pindel [25].
Like LUMPY, Manta [37] incorporates use of PE and SR methods. However, Manta also uses AS analysis. Manta overcomes the computational expense of AS methods by splitting the work into many smaller workflows which can be carried out in parallel. Manta scans the genome for SV and then scores, genotypes and filters the SV based on diploid germline and somatic biological models [37]. Manta can detect all structural variant types that are identifiable in the absence of copy number analysis and large-scale de-novo assembly, which is why this approach is also a good candidate for joint analysis of small sets of diploid individuals, tumor samples, and similar analyses. Both LUMPY and Manta are good at identifying SV break points with high resolution.
Many studies have been carried out to detect CNV using WGS data in various domesticated species: cattle [38], cats [39], chickens [40], dogs [41], etc. So far, there is no report of goat CNV discoveries using WGS data. The goal of this study was to identify CNV in the goat genome through the intersection of LUMPY and Manta outputs as a part of the characterization of African goats in conjunction with the ADAPTmap project [42]. Goats are a very important farm animal genetic resource for the livelihoods of African smallholders, and a deeper understanding of the goat genome is necessary to facilitate the improvement of goats in the region. This study aimed to generate a fine-scale CNV map for the goat genome.

Number and distribution of CNV
The number of CNV detected depended on the filter levels (low, medium, or stringent) and the cut-off point for CNV length (3 Mb or 10 Mb) as given in Supplementary Figure 11 (Additional file 2). Using precise SV only with moderate filters (PE + SR ≥ 5), LUMPY detected 8563 duplications and 230,497 deletions while Manta detected 24,088 duplications and 320,374 deletions. A combined data set with 244,876 deletions and 8677 duplications (totaling 253,553, translating into an average of 1393 CNV per animal) was derived from the intersection of the LUMPY and Manta sets after removal of variants shorter than 50 bp or longer than 3 Mb. The combined data set had more observations than the LUMPY data set (which had fewer raw CNV) because for some individuals, many short CNV from Manta intersected with few long CNV from LUMPY.
The CNV were distributed across the 29 autosomes as shown in Fig. 1. A vast majority of the CNV (96.6%) were losses. This is not unexpected, because all CNV detection methods suffer from an inherent deficiency in detecting insertions. In the case of CNV detection using WGS data, this limitation is even more pronounced with PE methods, because they detect insertions when the mapped reads are at a distance shorter than the fragment length, so they are not able to detect insertions larger than the insert size of the reference library [43]. This has also been supported by the observation that recall percentage is lower than 2 and 5% for medium (1-100 kb) and large (100 kb-1 Mb) duplications, respectively, for most of the SV-calling algorithms currently in use, including Manta and LUMPY used in this study [44].
Overall, the mean CNV length was about 3.3 kb, with a median of 1.3 kb. The distribution of the lengths of the CNV for each population are shown in Fig. 2 by CNV length category. A summary of the descriptive statistics of the CNV for the populations are given in Table 1. Most of the CNV losses (99.92%) were less than 100 kb long while 6.3% of CNV gains were longer than 100 kb. Despite the overwhelming proportion of losses over gains, there were more CNV gains observed over 100 kb than losses. Similarly, only 1.04% of the loss CNV were longer than 10 kb, while almost one-quarter (22.99%) of all gain CNV were over 10 kb. As a result, CNV gains were longer than CNV losses and had larger range in length. Deletions and duplications averaged about 2.3 and 31.5 kb long, with median lengths of 1.3 and 1.4 kb, respectively. There were no significant differences in the distribution of CNV across the five populations as shown in the percentile and sample QQ plots in Fig. 3.

Population CNV differentiation
Analysis of population differentiation (V ST ) as described by Redon et al. [11] showed that several CNV were highly differentiated between and across the populations. Some of these CNV overlapped with genes of importance in goats. Results for the pairwise population V ST tests and the V ST test across all the populations with their respective 99th percentile CNV V ST thresholds are given in Supplementary Table 1 (Additional file 1). V ST values for the pairwise tests are given in Supplementary Figures 1-10 (Additional file 2). The V ST values for genes that were in CNV that were highly differentiated across all populations are shown in Fig. 4. The gene DST was in a CNV with a very high V ST threshold across all the populations. DST has been associated with herpes virus and respiratory disease (BRD) in cattle [45]. Some CNV were highly differentiated both between and across populations. CNV with high differentiation between only some populations include the CNV corresponding to the genes BCO2, CCSER1 (FAM190A), COL24A1, CPNE4, CWC22, IMMP2L, KBTBD12, LAMA3, NAALADL2, RFX3, SEMA3D, SLC2A13, STPG2 (C4orf37), TAFA2 (FAM19A2), TMEM117, TMEM161B and VPS13B. The rest of the genes were in CNV that were highly differentiated across all populations.

Number and distribution of CNV regions (CNVR)
The lists of CNV regions (CNVR) by population are given in Supplementary Table 2 Table 3 (Additional file 1) while a distribution of CNVR by size and populations is given in Fig. 6. Over 92% of the CNVR were copy losses. There was a wide variation in the number and sizes of the CNVR between and among the populations. The fraction of copy gains or gains and losses was highest in the group of CNVR of at least 10 kbp, with 25% copy gains and 19% for losses/ gains (Fig. 6).

Functional annotation and gene enrichment analysis
Functional annotation was carried out for genes in global and private CNVR. Up to 2980 genes overlapped with the 6321 CNVR identified in this study. Up to 755 of these genes formed 24 clusters, with enrichment scores ranging from 0.0 to 1.89. Higher enrichment scores imply higher overrepresentation of the genes in the gene set for the gene enrichment term [46]. The top 3 clusters with the highest enrichment scores are given in Table 3 while the full list is given in Supplementary Table 5 (Additional file 1). The most significant GO terms identified in the analysis included retrograde endocannabinoid signaling; glutamatergic synapse; circadian entrainment; dopaminergic synapse; gastric acid secretion; long-term potentiation; salivary secretion; and calcium signaling pathway. CNVR private to populations and breeds overlapped with 172 and 620 genes, respectively. The GO terms associated with these genes based on functional analysis are listed in Supplementary Table 6 (Additional file 1). The genes that overlapped with the CNVR private to breeds were not significantly enriched in biological processes, molecular functions and cellular components, while the ones that overlapped with the CNVR private to populations were significantly enriched (P ≤ 0.05) with such terms as aldosterone synthesis and secretion; glucagon signaling pathway; insulin secretion; glutamatergic synapse; thyroid hormone synthesis; gastric acid secretion and phosphatidylinositol signaling system. The most common CNVR (chr6:115,822,332-115,825,687) includes the gene TMEM129 (transmembrane protein 129) that has been reported to be responsible for ubiquitination and proteasome-mediated degradation of misformed or unassembled proteins in the cytosol [47][48][49], and belongs to a network responsible for cellular assembly and organization, cellular function and maintenance, and cell cycle [50].

Discussion
This study identified CNV and CNVR in the goat genome using WGS data. Use of WGS for CNV detection is highly encouraged, because it overcomes many of the shortcomings of the other CNV detection methods such as the ones using array CGH and SNP data [19][20][21]. Genome-wide studies to discover CNV have already been done in other domesticated species, such as in Sus scrofa [51], Bos taurus [38,52] and Felis catus [39]. Here we provide a first glimpse of the goat genome CNV map at a dense genome coverage, using animals from 34 diverse breeds from the African continent. This addition is an important contribution, as goats are an important source of income and high-quality animal protein for small holder farmers in Africa. We used two software suites (LUMPY [30] and Manta [37]) for detecting SV to increase our confidence in the SV calls. Both software packages use split read and read-pair methods. They complement each other in that LUMPY makes use of read depth methods, while Manta draws heavily on genome assembly methods. Taking the intersection of SV calls from the two methods gives us confidence that the  number of false positives in the SV calls was kept to a minimum, although this means that some true SV were possibly filtered out. This study has shown that there are wide variations in the number and sizes of CNV in the goat genome between chromosomes, individuals and breeds. However, considering the small and variable numbers of samples within breeds, breed comparisons are not particularly meaningful. The results suggest that there are negligible differences in the sizes of CNV between populations. Some of the CNV displayed large differences between populations, suggestive of population-specific selective pressures.
A large proportion of the global CNVR identified in this study (65.1%) are within the CNVR reported by Liu et al. [18]. The remaining 34.9% may comprise false positive CNVR and CNVR that were missed by the PennCNV algorithm used in the other study, considering the limitation of CNV detection using SNP data, which include limited coverage for genome, low resolution, and difficulty in detecting novel and rare mutations. The CNVR coverage of 2.4% (59.2 Mb of about 2466 Mb of autosomal genome) found in this study is lower than the 4.8-9.5% SV coverage in the human genome [13], comparable to 55.6 Mb (2.0%) reported for cattle [38], later revised to 87.5 Mb (3.1%) [53].
V ST analysis showed that several CNV were highly differentiated among and across the populations. The genes in the highly differentiated CNV included BCO2  [54] reported that BCO2 is associated with the accumulation of carotenoids in the adipose tissue of sheep, leading to the yellow fat syndrome. The quality of semen (including total sperm motility, average path velocity and beat cross frequency) in Holstein-Friesian bulls has been associated with CCSER1 (FAM190A) as well as FAM155A [55]. GNRHR has been associated with number of days to first service after calving in dairy cattle [56] while IMMP2L is associated with cow conception rate [57]. The partial deletion of LAMA3 is responsible for epidermolysis bullosa in horses [58]; NAALADL2 is believed to be responsible for immune homeostasis [59], and TAFA2 (FAM19A2) is believed to be responsible for the regulation of feed intake and metabolic activities in mice [60]. Yamano et al. [61] reported that Fig. 6 Distribution of size of CNVR (in kbp) for each population. Orange is for copy gains and red is for CNVR with both copy gains and losses. The rest of the colours for copy loss for each of the five populations (magenta for Boer; blue is for the East African; green for Madagascar; brown for Southern African and purple for West African) TOMM70 is responsible for integral mitochondrion proteins and for metabolism. Functional annotation and clustering analysis revealed that the CNVR identified in the study have genes that are significantly enriched with many biological processes, molecular functions and cellular components, some of the most significant of which are retrograde endocannabinoid signaling, circadian entrainment and long-term potentiation. The retrograde endocannabinoid signaling system is a complex and diverse regulator of synaptic function [62], and is responsible for many diseases in the nervous system and peripheral organs. In the human genome, this system is widely considered as a potential target for treating conditions such as alcoholism [63]. A CNVR in the cannabinoid receptor 2 (CNR2) region has been reported in the human genome, but its effect has not been fully characterized [64]. Zajkowska et al. [65] suggested that there is need to explore genetic variation in the system from the perspective of copy number of variations.
Circadian entrainment is an important aspect of animal behavior and adaptation, especially considering the wide range of environmental conditions the animals are exposed to. An example of goat adaptation to the environment is their ability to rapidly change the size of their foreguts in response to changes in the environment [66]. Goats tend to be active during some parts of the day only [67], and this varies with season [67], suggesting a considerable amount of circadian entrainment. The increased importance of the biological process "response to stimulus" (GO:0050896) in the highly differentiated CNV may also support the hypothesis of the importance of circadian entrainment in goats.

Conclusions
This study presents the first fine CNV map of the African goats based on WGS data. This information will prove invaluable for further improvement of goats, especially on African continent, as more phenotype data becomes available, through CNV or CNVR association analyses and other approaches.

Sample description
The data used in this study was generated from 182 goats representing 34 breeds from 9 Sub-Saharan African countries (Ethiopia, Kenya, Madagascar, Malawi, Mali, Mozambique, Tanzania, Uganda, and Zimbabwe), and these countries were grouped into four populations based on geographic locations and a fifth population of Boer goats obtained in Tanzania and Zimbabwe. The Boer goat is a special breed widely used in Africa and much of the world [68]. The samples were previously genotyped using the Illumina Goat SNP50 BeadChip [69] as described by Bertolini et al. [70], Cardoso et al. [71] Fig. 7 Location of the global CNVR across the 29 autosomes. Blue is for loss; red is for gain and green is for both loss and gain and Colli et al. [72], and some of them were also used for detection of CNV using 50 K SNP chip data, as reported by Liu et al. [18]. A list of the breeds, populations and samples sizes used in the analysis is given in Table 4.
Sample processing was done by Edinburgh Genomics using the Edinburgh Clinical Genomics method. This approach uses Illumina SeqLab products and services, including, Illumina TruSeq library preparation, Illumina cBot2 cluster generation, Illumina HiSeqX sequencing, Hamilton Microlab STAR integrative automation, and Genologics Clarity LIMS X Edition as outlined in Supplementary Table 7 (Additional file 1). Quality control information for the samples is given in Supplementary  Table 8 (Additional file 1).

Detection of SV
SV were detected using LUMPY version 0.2.13-85-gc1bcea1 and Manta version 1.5.1, which are two of the most used algorithms for detecting SV. In LUMPY, the "lumpyexpress" script was used. This script runs automated breakpoint detection for standard analyses. It uses SAMBLASTER [76] to extract split and discordant reads from BWA-MEM-aligned Binary Sequence Alignment Map (BAM) files. Default options were used, including minimum non-overlap and minimum sample weight set to 20 and 4, respectively. In Manta, the "configManta.py" script was used to process each sample, with default options including minimum variant candidate size (8); minimum candidate spanning count (3); minimum scored variant size (50); minimum diploid variant score (10); minimum diploid variant score pass point (20); minimum somatic score (10); and minimum somatic score pass point (30). The "runWorkflow.py" scripts were run in parallel to extract the SV for each sample.
Post-processing of SV SV from LUMPY were genotyped with svtyper version 0.6.1 [77], which uses a Bayesian maximum likelihood algorithm to determine the most likely genotype of each base-pair. Variant call format (VCF) files from the two Fig. 8 Distribution of the CNVR. a, b Number of CNVR found in different numbers of breeds and populations, respectively. c, d Distribution of CNVR found in only a single breed and only a single population only, respectively. In C, only 30 breeds had private CNVR software packages were converted to browser extensible data (BED) format for downstream analysis using svtools version 0.5.0 [78]. Various levels of SV post-processing parameters were used to come up with the CNV calls from the SV calls. The parameters included: 1) precision of SV calls (whether imprecise SV were included in computation of the CNV calls); 2) point of application of the lower SV length cut-off point (before or after merging Manta and LUMPY SV); 3) stringency of the SV call filters (low, medium, and high stringency); and 4) upper SV length cut-off (3 or 10 Mb). Stringency of SV call filters was in terms of the number of PE and SR required as evidence supporting an SV. Consensus SV were obtained by identifying the intersection of the SV from LUMPY and Manta using BEDTools version 2.26.0 [79] with default settings.

Derivation of copy number variations
CNV were defined as SV duplications and deletions longer than 50 bp [80]. SV longer that 3 Mb were also filtered out, because putative CNV in the goat genome are usually much shorter than this length. Visualization of the SV was done using R [81] package circlize version 0.4.7 [82].

Population CNV differentiation
A measure of population differentiation (V ST ) as described by Redon et al. [11] was computed based on normalized read count values for each CNV, similar to the method used in PECNV as described by Liu et al. [83], which was in turn based on clustering algorithms described by Cridland et al. [84] and transposable element detection algorithms described by Rogers et al. [85].
Read count values were corrected for size of the consensus CNV, batch effect, variable GC content and genomic mappability as described by Liu et al. [83]. Regional and batch effect correction was done by computing reads per kb per million mapped reads (RPKM) as described by Mortazavi et al. [86], where RPKM ¼ 10 9 ÃRC TRCÃS , where RC is the read count of a region, S is the size of the region and TRC is the total number of mapped reads in the library. GC content and mappability correction was done on the RPKM using the formula used by Yoon et al. [29], where adjusted read count is given by RPKMÃm m GC where m GC is the median GC content of all regions with the same read count and m is the median GC of all regions. This approach is similar to the read depth approaches used in CNVnator [28] and in CNVcaller [87]. The normalized read count values were treated as proxies of log R ratio (LRR) values normally obtained from array analysis. As defined by Redon et al. [11], V ST was computed as V T −V S V T , where V T is the variance in LRR among all unrelated individuals and V S is the average variance in LRR within each population. CNV V ST testing was done pairwise (for each combination of two populations) and (separately) across all the 5 populations. CNV with V ST values above the 99th percentile of all V ST values for each comparison were treated as being highly differentiated. We searched for these highly differentiated CNV in the Golden Helix Genome Browse® software (version 3.0.0) (https://www.goldenhelix.com/) using the ARS1 caprine genome reference assembly to identify the genes in the CNV.

Determination of CNV regions
CNV regions (CNVR) were obtained by merging CNV that overlapped by at least 1 bp within populations (population CNVR) and across all the individuals (global CNVR) using the "merge" function in BEDTools version 2.26.0 [79].

CNVR functional annotation and gene enrichment analysis
A list of genes for the goat genome was downloaded from the NCBI website (https://www.ncbi.nlm.nih.gov/ gene). The Database for Annotation, Visualization, and Integrated Discovery (DAVID) Bioinformatics Resources (version 6.8) [88][89][90] was used to identify if genes in the CNVR have significant biological, cellular or molecular function. Functional analysis was done using default parameters, with significance of enriched terms determined at P ≤ 0.05. Further information about various genes was obtained from the GeneCards (www.genecards.org) database.