Introduction

Structural variants (SVs) are generally characterized as genomic alterations exceeding 50 bp and encompass various forms, such as deletions, duplications, insertions, inversions, and translocations1. Due to their size, SVs are more likely to influence genetic variability than single nucleotide polymorphisms (SNPs) or small insertion-deletions (InDels)1,2,3. In contrast to SNPs and InDels, the development of SVs detection pipelines presents greater challenges and requires further optimization. This complexity arises from factors such as the types of reads, the range of sizes, the methodologies employed, and the presence of multiple SVs at identical locations 1,4.

Many studies have used short-read (SR) sequences to detect SVs, aligning them to a reference genome and using classical detections tools 5,6,7,8. These SR-based tools often rely on alignment-related criteria, for instance on read-depth, split-reads, or abnormal distance between paired reads, from the alignment to detect SVs 9,10,11. Recent studies have indicated that this strategy has limitations in accurately identifying SVs 12,13. The limitations are attributed to the SR lengths and technical artifacts associated with SR alignment, especially in regions with repetitive sequences or high GC content 12,13. For example, half of deletions and three quarters of insertions cannot be detected using SR alone, even with a combination of multiple variant callers 13. Therefore, the use of long-read (LR) sequences is advisable, as they allow for a more comprehensive SVs detection 14,15.

LR-based SVs detection tools are mainly categorized into assembly and alignment based 15. Briefly, the former assembles the reads de novo, and then compares this assembly to the reference genome to call SVs. In the alignment approach, LR sequences are mapped directly against the reference genome and SVs are detected according to different alignment signatures to the reference genome 4,16. It has been suggested that alignment-based methods generally have better genotyping accuracy than assembly-based methods, even with lower coverage 4,15,16. Moreover, the alignment-based methods allow, with no additional work, to detect heterozygosity whereas assembly based requires haplotype resolved assemblies to detect heterozygosity, which is not a trivial task 17,18.

Several LR sequencing projects are underway, supported by the ever decreasing cost, creating an ideal situation where SVs will be discovered from LR data 4,19,20. Currently, the main LR platforms are PacBio single molecule real-time (SMRT) and Oxford Nanopore Technologies (ONT) which produce reads length ranging from ten to several hundred kilobases 13. In addition, PacBio protocols are categorized into continuous-long-reads (CLR; > 30 kb and one polymerase pass) and high-fidelity (HiFi; 10–30 kb and multiple polymerases passes) based on the library insert size and number of polymerase passes over the fragment. This data type has emerged as the most prevalent in contemporary usage, owing to the benefits of both extended range and enhanced base accuracy (> 99%) 13,20.

Nonetheless, SR platforms account for a significant portion of genomic research, sequencing millions of genomes annually across various populations and species. This is primarily attributed to the reduced cost per base and the simplified library preparation process 4,14. Thus, SVs detection from SR might help us to better understand the causality of Mendelian and/or complex traits of interest. The challenge is how to make the best use of SR data given these discrepancies.

In addition to classical SVs callers from SR, there are recent tools available for genotyping known SVs using SR, such as GRAPHTYPER 21, PARAGRAPH 22, SVTYPER 23, and Variation Graph toolkit (VG) 24. The different variants at a given locus are embedded in a graph, each sub-path corresponding to a given allele. SR sequences are then re-aligned to this graph, and alignment evidences at and around the breakpoints are used to genotype the SVs 22,25. Using simulated data, previous studies have benchmarked the performance of various SR-based SVs callers and genotypers with a presumably known truth-set of SVs 26,27. Nonetheless, the evaluation of SVs detection using current datasets, especially in the context of livestock, remains quite restricted 28.

Sequencing followed by alignment thousands of recognised cattle breeds to the current bovine reference genome of ARS_UCD1.2 may lead to discovery of a vast number of common and rare SVs 19,29. Recently, the Bovine Pangenome Consortium has commenced the international effort aiming, among others, to capture these SVs diversities from global cattle populations via pangenome approach 19. In the past, a work on the detection of 6,462 SVs in three major French breeds was carried out by alignment of SR data of 62 bulls 5. Using SR data, another study detected approximately 68,000 SVs from 310 Holstein cattle of which approximately 10% overlap with the current BovineHD SNP array 30. An investigation in Holstein and Nordic Red Dairy cattle has associated three and twelve SVs toward several fertility traits, respectively 31. In addition, a single deletion was reported with an antagonistic effect on fertility and milk production traits in Nordic Red Dairy cattle 32. Identified SVs in cattle were also found to be associated with gene expression, such as SLC13A4 and TTC7B in testis tissue 33. In a similar manner, various SVs were notably linked to gene expression, which leads to phenotypic diversity associated with high altitude and coat color in domestic yak 34.

In light of these circumstances, the aims of the present study are to explore and benchmark the outcomes of SVs detections from various approaches. This includes (i) demonstrating the impact of different LR technologies and callers on SVs detection; (ii) highlighting the performance of different SVs callers and genotypers using SR information; (iii) exploring the effect of incrementing the number of reference samples and of applying different thresholds on the performance of the genotyping tools; and finally, (iv) using the information of genotyped SVs for population-based analysis in a larger SR dataset of French cattle.

To date, this study represents the most thorough examination of SVs, utilizing extensive LR and SR data derived from current cattle datasets. Consequently, our findings illustrate the potential practical uses of SVs detection and genotyping within the bovine genome, and possibly in other species, to further investigate its relationship with the genetic variability of significant Mendelian and complex traits.

Results

In this study, we used TRUVARI 35 to benchmark two sets of SVs, designated as the base set (putatively true) and the comp set (inquired), respectively. In the benchmarking, this tool considers types of SVs, reference distance, reciprocal overlap, size similarity, sequence similarity, and genotype matching between the two sets, outputting several metrics of precision, accuracy, F1 score and genotype concordance scores (see “Methods”). The assignments of base and comp sets correspond to the benchmark design in each of the following analysis.

Evaluation of SVs detection tools on different LR data

A total of 176 samples were LR sequenced. Except one, all samples were LR sequenced either with PacBio HiFi, ONT, or PacBio CLR, with the predominant technology being CLR (Table 1; Table S1). A single Charolais heifer (CHR18) was sequenced with three LR technologies. Given this dataset, it was a requisite to identify the most reliable tool that can consistently detect SVs, particularly on deletions and insertions, across all three LR sequencing technologies.

Table 1 Descriptive statistics of of long-read (LR) and short-read (SR) samples used in this study.
Full size table

For this, we took advantage of CHR18 sequenced with three LR technologies (Table S1), with coverage up to 40 × for HiFi, 52 × for ONT and 42 × for CLR 36. We deemed these coverage levels adequate for a comprehensive analysis in light of the criteria established by earlier studies for assembly and variant discovery 18,20. After the alignment of reads from each technology to the ARS_UCS1.2 reference genome, SVs were detected using CUTESV 37, PBSV 38, and SNIFFLES 39, commonly-used software tools.

First, we characterized the SVs detection results on HiFi data. After quality control, CUTESV, PBSV, and SNIFFLES detected 11,002, 11,575, and 13,567 deletions, respectively (Fig. S1). We conducted TRUVARI 35 benchmarking to identify intersections of SVs detected by the tools. For insertions, CUTESV, PBSV, and SNIFFLES detected 11,759, 10,258, and 13,913, respectively. Noteworthy, 10,000 deletions and 8866 insertions, were commonly detected by the three tools (Fig. 1a). We also observed that SNIFFLES identified 4328 specific SVs, which was significantly higher than those detected by CUTESV (784) and PBSV (566).

Fig. 1
figure 1

Properties of structural variants (SVs) discovered by various long-read (LR) technologies with different software on the genome of a single Charolais heifer (CHA18). (A) Venn diagram indicating the number of overlapping SVs called by CUTESV, PBSV and SNIFFLES tools on the heifer’s PacBio-HiFi data. (B) Plot of total number SVs discovered by the three callers against chromosome size. (C) Frequency of discovered SVs according to their length (x-axis) for deletions (left) and insertions (right) using the three tools. The first, second, and third rows indicate the LR sequencing of Pacbio-HiFi, Oxford-ONT, and PacBio-CLR, respectively. (D) F1 score output from the TRUVARI bench. The base set is the SVs called from HiFi and the comp set is the SVs called from ONT or CLR data using the CUTESV, PBSV and SNIFFLES callers.

Full size image

The total numbers of SVs called by all three tools were highly correlated (R coefficient ≥ 0.85, Spearman’s correlation test) with chromosome size (Fig. 1b). Heterozygous [0/1] and homozygous alternative [1/1] SVs were observed for this heifer called by the three tools. Although the fraction is relatively small, SNIFFLES also reported 99 insertions with uncalled genotype [./.] and 1764 deletions and 1827 insertions with homozygous reference [0/0] (Fig. S1; Table S2).

A symmetric distribution of SVs based on their length, between deletions and insertions, is expected 20,36,40. For the HiFi data, SVs distribution followed this pattern between deletions and insertions across the three callers (Fig. 1c). We observed a trend where the number of detected SVs declined as variant size increased. Clear peaks around 143 bp, putatively correspond to SVs due to transposable elements families (SINE, LTR, LINE) 36, were detected for deletions (345, 287, and 294 for CUTESV, PBSV and SNIFFLES, respectively) and for insertions (256, 225, and 284 for CUTESV, PBSV, and SNIFFLES, respectively) (Fig. 1c; Fig. S2). Symmetric distributions based of lengths between deletions and insertions were similarly detected by the three tools in other HiFi-sequenced samples, i.e., Abondance (ABO1), Tarentaise (TAR1), and Vosgienne (VOS1) (Figs. S3–S5).

For ONT data, the distribution pattern remained similar, but the peaks were lower and became less pronounced, especially for insertions, compared to HiFi data. For CLR data, the deletions length distribution mirrored that of HiFi and ONT data. However, the distribution pattern for insertions changed significantly, with an inflation of insertions around 50 to 100 bp, detected by CUTESV and SNIFFLES. In contrast, PBSV consistently detected the peak around 143 bp for both deletions and insertions across HiFi, ONT, and CLR technologies (Fig. 1c; Fig. S2). Notably, this inflation of insertions was found not only in the Charolais heifer, but also in almost all individuals sequenced by PacBio CLR, for example in Holstein (HOL28), Montbéliarde (MON16), and Normande (NMD19) (Figs. S6–S8).

To further evaluate the performance of CUTESV, PBSV and SNIFFLES across different LR data, we performed a benchmark between SVs called from ONT or CLR data (as the comp set) with the base set, SVs called from HiFi data. PBSV outperformed CUTESV and SNIFFLES with F1 scores above 0.91 for deletions and above 0.84 for insertions in both the ONT and CLR datasets (Fig. 1d; Table S3). This indicates that PBSV calls showed the highest consistency in SV calling across HiFi, ONT and CLR data.

We further evaluated whether an ensemble calling, enriching the SVs discovered by PBSV with SVs called by CUTESV and/or SNIFFLES, could improve the benchmarking results. In this study, we utilized both the intersect and union approaches for the pairs of PBSV-CUTESV, PBSV-SNIFFLES, and a combination of three callers, applying them to CLR data as the comparison set in relation to the base set (see “Methods”). Our analysis revealed that none of the combinations achieved an F1 score that surpassed the default PBSV on CLR vs. PBSV on HiFi comparison (0.91 for deletions and 0.84 for insertions) (Table S4). Thus, we did not identify any further benefits in enriching SVs discovered by PBSV on CLR data with SV calls from other tools.

Of note, PBSV recommends using additional information from tandem repeat regions for increasing sensitivity and recall of SVs. We annotated tandem repeats and used this annotation during SVs calling in three HiFi samples of ABO1, TAR1, and VOS1 (Table S5). We compared the results to SVs previously discovered without annotation. The average F1 score was 0.98 corresponding to a relatively low number of SVs (~ 300) that were different without tandem repeat annotation (Table S5). Consequently, considering the balance between the potential gains from SVs and the computational resources needed, we proceeded with the analysis using SVs identified by PBSV on LR data, excluding tandem repeat annotations.

At this stage, considering the benchmarking results for the Charolais heifer and the additional CLR individuals available in our dataset (Table S1), we used SVs detected by PBSV for the subsequent analysis in this study regardless of whether the sample was sequenced with HiFi, ONT or CLR.

Evaluation of SR-based SVs genotyping tools

We evaluated the feasibility of using SR data to recover the genotypes of identified SVs that were originally discovered through LR data. The analysis included three main groups of tools based on read alignment and requirement of known SVs input information: (1) DELLY 10, LUMPY 11, and MANTA 9 which use evidences from the BAM file alignment to call SVs; (2) GRAPHTYPER 21, PARAGRAPH 22, and SVTYPER 23 which use the BAM file alignment and known SVs for local re-alignment nearby breakpoints to genotype SVs; (3) Variation Graph toolkit (VG) 24 which creates a genome-wide variation graph from SVs and aligns the entire reads to the graph for genotyping the SVs (Fig. 2a). Here, TRUVARI benchmarked between comp sets, SVs that were genotyped using various tools with SR data, and the base set, SVs called by PBSV from CLR data (Fig. 2a). We used both F1 score and genotype concordance score produced by TRUVARI to measure the ability of the tool to correctly detect SVs with correct genotype.

Fig. 2
figure 2

Structural variant (SV) gGenotyping of known SVs from short-reads (SR) data. (A) Workflow from benchmarking SVs genotyping tools. (B) Performance of SVs genotyping tools on CHA18 for deletions (left) and insertions (right) using F1 indicator and genotype concordance scores at different levels of SR coverage (circle size).

Full size image

Again, we used the CHA18, for which both LR and SR are available at different levels of coverage (see “Methods”). The F1 scores of the genotyping tool were higher when using SVs information (Fig. 2b). For example, on deletions, the F1 score of DELLY and LUMPY were both 0.63, comparable to GRAPHTYPER at the highest coverage of 60x. In contrast, MANTA’s score was 0.79, while genotyping tools using input SVs information, such as VG, PARAGRAPH and SVTYPER, achieved F1 scores higher than 0.94. Note that LUMPY only reports SV breakpoints without genotype information. Only MANTA, GRAPHTYPER, SVTYPER and VG genotype insertions. Their F1 scores for insertions were lower compared to deletions, 0.48 for MANTA and 0.47 for GRAPHTYPER. Meanwhile, PARAGRAPH and VG slightly increased their F1 scores from 0.98 to 0.99 and 0.94 to 0.96, for deletions and insertions, respectively.

The performances of the genotyping tools were predominantly influenced by the coverage, with the notable exceptions of GRAPHTYPER and VG. For both deletions and insertions, the coverage affected the F1 scores of DELLY, LUMPY and MANTA. The genotype concordance of paragraph and SVTYPER were also affected by the coverage. Furthermore, considering that the SR alignments are carried out independently for each SVs genotyping tool, the resource requirement is better optimized for VG because it only used one third of the time (92 cpu-hours) compared to the other genotyping tools (330 cpu-hours) for the SR alignment with 60 × coverage (Table S6).

We expanded the evaluation of SVs genotyping tools by conducting an independent analysis of 154 samples with both LR and SR available (Table 1, Table S1) comprising 14 different bovine breeds. Here, TRUVARI benchmarked between comp sets, genotyped SVs using various tools with SR data, and the base set, consisted of SVs called by PBSV from CLR data (Fig. 2a). After removing the outliers, 148 samples were retained in the analysis. In this group of individuals, a greater number of deletions were identified compared to insertions (Fig. 3a). The number of SVs differed significantly, ranging from 15,472 in HOL19 to 23,086 in AUB3.

Fig. 3
figure 3

Structural variants (SVs) genotyping on 148 individuals with long and short-reads available. (A) Histogram indicating the number of SVs in the base set called by PBSV on CLR data. (B) Performance of SVs genotyping tools on deletions (left) and insertions (right) based on F1 and genotype concordance scores. (C) Boxplot with Wilcoxon rank test on the number of SVs genotyped exclusively by PARAGRAPH and VG. (D) Recall and genotype concordance scores of commonly (intersect) and exclusively genotyped SVs by PARAGRAPH and VG against the base set of the respective individuals.

Full size image

Performances of the genotyping tools were consistent with observations made on CHA18 (Fig. 3b). Tools were grouped similarly based on both F1 and genotype concordance scores, reflecting coherent clustering patterns. For F1 scores, PARAGRAPH, SVTYPER, and VG demonstrated significantly superior performance in genotyping deletions compared to the other tools (Fig. 3b; Fig. S9). For insertions, PARAGRAPH and VG showed significantly better performance compared to the other tools (Fig. S10).

We conducted an additional comparison of SVs identified by both PARAGRAPH and VG, leading to the classification of SVs into four distinct sets for all individuals: intersect SVs genotyped by VG, intersect SVs genotyped by PARAGRAPH, exclusive SVs genotyped by VG, and exclusive SVs genotyped by PARAGRAPH. Each SV set was subsequently compared to the base set of the respective individual. It is noteworthy that there were as many as five times more SVs uniquely genotyped by PARAGRAPH compared to VG (Fig. 3c). When categorizing the genotyped SVs into groups, the precision and recall rates of each group became less informative due to their significantly reduced numbers in comparison to the base set. Thus, the primary objective of this comparison was to determine the accurate assignment of heterozygous [0/1] or homozygous alternative [1/1] genotypes as established in the base set. Our analysis revealed that VG exhibited superior genotype concordance compared to PARAGRAPH for both intersecting and exclusive SVs, encompassing deletions and insertions (Fig. 3d). Across the 148 samples, the average genotype concordances scores were 0.78 for PARAGRAPH and 0.89 for VG.

In light of the independent TRUVARI benchmarking results on the CHA18 and other 148 individuals between their LR-based SVs and SR-based genotyped SVs, we have determined that VG is the most performing genotyping tool for recovering SVs from SR. Thus, VG will be utilized in the forthcoming analyses.

Extending the number and diversity of reference SV panels

In the previous section, we assessed the performance of SVs genotyping tools on a sample-by-sample basis, with VG achieving the best results. In this section, we investigated the influence of gradually incorporating samples from various breeds into the variation graph on the performance of SVs genotyping by VG. The workflow (Fig. S11) began with merging the base sets of reference samples using JASMINE 41, followed by the construction of the variation graph (see “Methods”). Validation samples were then genotyped using SR mapped against the variation graph, and the resulting genotyped SVs were compared to the base set of each respective validation individual.

We selected validation samples consisted of two individuals each from HOL, MON, and NMD. We created the first reference panel by merging the base sets (SVs discovered by PBSV on CLR data) of 23 HOL samples. The second panel was created by adding the base sets of 23 MON reference samples. The third panel was created by adding the base set of 24 NMD reference individuals. And the last panel consisted of all reference samples, resulting in a reference panel of 145 samples from 14 breeds. Base sets of validation samples were not included in the creation of any of these panels. As the number of reference samples increased, the total number of SVs with distinct alleles and their corresponding bubbles in the graph also increased (Table S7). Across the four panels, 94.4% of the bubbles created were biallelic, while the remaining bubbles were multi-allelic resulting from SVs with identical breakpoints but multiple alternative alleles. These multi-allelic bubbles were already present in the single-breed panel but became increasingly abundant as additional reference breeds were added (Table S8). For example, a bubble indicating an insertion at chr 25: 39,613,777 initially appeared as biallelic, with one alternative insertion allele in panel one (HOL only panel). This remained unchanged in panel two (HOL + MON panel). However, in panel three (HOL + MON + NMD panel), the bubble became more complex by acquiring an additional allele common to NMD samples. By panel 14, which included 14 breeds as references, this bubble had three alternative alleles (Fig. 4a). Overall, the number of bubbles almost doubled from panel one to panel 14, reflecting the increased diversity introduced by the broader breed representation (Fig. 4b; Table S7).

Fig. 4
figure 4

Increasing the number of samples in the SVs reference panels affect the genotyping performance. (A) The variation graph bubble indicates the insertion node at position 25: 39,613,777 created from (in clockwise): P1 includes 23 HOL samples; P2 includes 46 HOL and MON samples; P3 includes 70 HOL, MON, and NMD samples; P14 includes 145 samples from 14 breeds. (B) Increasing number of bubbles in the variation graph as the number of reference samples increases. (C) Benchmarking of genotyped SVs of 6 validation individuals from SR data on different variation graphs. Left and right columns represent deletions and insertions, respectively. The values in the graph dots indicate the number of breeds on which the variation graph was created. Breed abbreviations are: HOL (Holstein), MON (Montbéliarde), NMD (Normande). The other 11 breeds are Abondance, Aubrac, Blonde d’Aquitane, Brown Swiss, Charolais, Limousine, Rouge Flamande, Parthenaise, Simmental, Tarentaise, and Vosgienne.

Full size image

The precision and recall rates of the validation samples were shifted depending on the reference panels used to construct the variation graphs (Fig. 4c). In panel one, the two HOL samples achieved the highest recall rates, approximately 0.82 for deletions and 0.84 for insertions. The inclusion of 23 MON into the reference for the creation of panel two improved the recall rates of two MON samples (Fig. 4c; Table S8). For example, MON1’s recall rate for deletions improved from 0.73 in panel one to about 0.83 in panel two. Similarly, the NMD sample’s recall rates also improved from around 0.73 to 0.78 for deletions and from around 0.72 to 0.78 for insertions in panel two. When NMD samples were then added into the reference to form panel three, their recall rate further improved compared to when this breed was excluded. Specifically, the recall rates of two NMD samples reached 0.83 for the deletions and 0.84 for insertions in panel three. As more breeds were added into the reference panel, recall rates slightly increased and became more consistent across validation samples. For example, recall rates for the HOL, MON, and NMD validation samples on panel 14 were around 0.83 for both deletions and insertions.

On the other hand, precision rates tended to decline as more breeds were added to the reference panel. This pattern was observed uniformly among all validation individuals, showing an average decrease of 0.08 in precision rate from panel one to panel 14. For example, HOL1 had a precision rate of 0.77 on panel one and only 0.67 on panel 14. As a side note, no clear trend was observed in genotype concordance rates as the reference panels evolved (Fig. S12).

We deduced that incorporating individuals from particular breeds into the reference panel facilitates the genotyping of breed-specific SVs, thereby improving recall rate. Nevertheless, as the number of breeds in the panel increases, VG was compelled to genotype SVs that were absent from the validation samples, leading to a reduction in precision.

Optimum parameters for building a reference panel of SVs across 14 breeds

Here, we used the same set of reference and validation samples as described in the previous section. We merged SVs from 145 samples, resulting in 107,274 and 126,008 SVs with distinct alleles using the default and adjusted parameters of JASMINE, respectively (see “Methods”). Subsequently we implemented eight VARCALLS which represented the number of variant calls that supported the merged SVs, of 2, 3, 5, 7, 13, 20, 27, and 36 as filters to the two merged SV sets. This process resulted in the creation of 16 reference panels. These SV panels served as the basis of constructing variation graphs on which SR of validation samples were then aligned.

F1 scores increased linearly with the application of more stringent VARCALLS thresholds (Fig. S13). Higher VARCALLS thresholds reduced the number of SVs to be genotyped, thereby increasing the precision rate (Table S10). In contrast to precision, recall rates across validation individuals peaked between thresholds of five and seven, when deletions and insertions were considered together (Fig. 5a). By separating deletions and insertions, the optimal VARCALLS threshold was determined to be between seven and thirteen (Fig. S14). Therefore, we selected a VARCALLS threshold of seven as the optimal criteria for filtering. As a result, only SVs detected in at least seven samples were retained in the reference panel.

Fig. 5
figure 5

Finding optimal parameters for creating a reference panel of SVs from 14 breeds. (A) Benchmarking of genotyped SVs from 6 validation individuals on the reference panel created from 145 samples from 14 breeds. JASMINE SVs merging was applied with default and adjusted parameters combined with different VARCALLS thresholds of 2, 3, 5, 7, 13, 20, 27, 36 as indicated by the points in the graph. Square and round dots indicate default and adjusted JASMINE merging parameters. (B) Number of input SVs not translated into bubbles or biallelic bubbles in the variation graphs across VARCALLS threshold.

Full size image

At the same VARCALLS threshold, employing the adjusted merging parameter led to an increase in the number of SVs in the reference panel when compared to the default parameters (Table S9). However, the difference between the numbers of input SVs and bubbles were significantly higher with the adjusted merging parameters of JASMINE in the comparison to the default (Fig. 5b). Furthermore, within a VARCALLS of seven, the number of biallelic bubbles in the variation graph of the default merging was higher (52,053) than with the adjusted parameters (51,576) (Table S9).

Taking this into consideration, we sought to identify the optimal parameters to achieve an optimal balance between the recall rate and the total number of SVs within the fourteen-breed’s panel. Based on our evaluation of the validation samples, we considered that merging with the default JASMINE parameters and applying VARCALLS threshold of seven provided a satisfactory balance for creating the all-breed panel when merging from the base set of all LR samples.

Genotyping SVs in a large cohort of SR samples

Considering the recommendations from the previous section, we merged the base sets of 176 samples from 14 breeds (Table 1) using the default JASMINE parameters, resulting in 159,792 SVs. After applying the VARCALLS threshold of seven, we retained 58,427 SVs. This filtered set was then used as input to construct the variation graph, which contained 55,309 biallelic bubbles representing 25,191 deletions and 30,118 insertions (Table S10). In total, these SVs provided alternative of 26,538,017 bp through deletions and 25,495,462 bp through insertions to the bovine reference genome.

The length distribution of SVs in the variation graph was similar for deletions and insertions, indicating a high-quality SV panel (Fig. 6a). However, we observed that deletions tended to be longer than insertions. Despite this, the graph contained a higher number of insertions than deletions (Table S10). Based on this graph, we then performed SVs genotyping for 571 SR samples (Table 1). Across individuals, the average number of genotypes were 377 uncalled [./.], 32,645 homozygous-reference [0/0], 14,535 heterozygous [0/1], and 7,750 homozygous-alternative [1/1] (Table S11).

Fig. 6
figure 6

Genotyping of structural variations (SVs) in 571 short-read (SR) samples using a variation graph constructed from 176 long-read (LR) samples. (A) Length distribution of genotyped SVs from the variation graph, showing deletions (left) and insertions (right). (B) Annotation of genotyped SVs based on their genomic region type. (C) Recall rates of genotyped SVs for 148 samples with both LR and SR data. (D) PCA of genotyped SVs across 571 SR samples. (E) Upset plot of major SVs (allele frequency > 70%) across all 14 breeds and 6 breeds with more than 30 SR samples. The inset displays SVs allele frequencies across 571 samples, with a vertical line indicating the threshold for major SVs. Breed abbreviations: AUB (Aubrac), BAQ (Blonde d’Aquitaine), CHA (Charolais), HOL (Holstein), LIM (Limousine), MON (Montbéliarde), NMD (Normande).

Full size image

Most of genotyped SVs were located in intergenic (44.13%) and intronic (44.42%) regions, while 5.01% were upstream and 4.97% downstream of known genes (Fig. 6b). Only a small fraction (0.69%) was found within exons. Using snpEff 42, 86,692 effects were associated with these SVs, with 98.86% classified as having a modifier effect and less than 1% categorized as high impact (Fig. S15). These high impact SVs comprised of 370 deletions and 62 insertions overlapping with 368 protein coding genes and 55 RNA-related regions. On average, a genotyped SV was observed approximately every 47 kb.

On 148 samples with both LR and SR available, we benchmarked their genotyped SVs against their base sets from LR. For each breed, the average recall rates were higher for insertions compared to deletions (Fig. 6c). For instance, the difference between deletion and insertion recall rates were around 0.03 across ABO, HOL, MON, and NMD. No clear difference in recall rates was observed between the breeds due to different size of reference or genotyped samples. For instance, the average recall rate of insertions on ABO was 0.922 while HOL, MON, and NMD had an average recall rate of 0.923, 0.920, and 0.921, respectively.

Average recall rates of breeds with reference samples below seven, corresponding to VARCALLS filter of seven, were among the lowest, such as the BSW (n = 5), PAR (n = 3), and RDC (n = 3). Yet, recall rates of SIM (n = 5), TAR (n = 6), and VOS (n = 5) were comparable to breeds with higher number of reference samples, for instances to LIM (n = 10) and CHA (n = 18). This suggests that most of breed-specific SVs belonging to BSW, PAR, and RDC were removed during the filtration step. In contrary, SVs found within SIM, TAR, and VOS were well-represented from the merging of 176 reference samples with the applied filter criteria.

Principal component analysis (PCA) of the genotyped SVs revealed a distinct differentiation among HOL, MON, and NMD when considering the first two components (Fig. 6d). The first component explaining 34.26% of the variance mainly separated HOL from the other French breeds. The second component, accounting for 18.23% of the variance, further separated MON and NMD at opposite extremes, while the remaining breeds closely clustered. Additionally, the fourth principal component, explaining 6.86% of the variance, further differentiated CHA, MON, NMD, and LIM. This result was consistent with the PCA performed on the SVs detected from 176 LR samples used as the reference for constructing the variation graph (Fig. S16).

Across 571 samples, the majority of SVs were found to be rare, with their frequencies primary clustered around 4–5% (Fig. 6e). The frequency distribution of genotyped SVs revealed two distinct extremes, i.e., SVs fixed with homozygous-reference or homozygous-alternative across all samples. At the first end of SVs with homozygous-reference, there were 339 deletions and 296 insertions detected, indicating that no SVs alleles were genotyped. Out of these SVs, seven were derived from samples with only LR sequences, while the remaining SVs were initially detected in the base set of 148 samples that have both LR and SR data (Figure S17). Of the 635 SVs in this category, 332 (52.28%) overlapped with known repeat regions in the bovine reference genome, including LINE (36.73%), SINE (35.10%), LTR (10.91%), simple repeat (7.55%), and DNA transposons (5.10%). The GC contents of the reference (for deletions) or alternative (for insertions) sequences were significantly higher with a p-value < 2.2e−16 (unpaired Wilcoxon-rank test), compared to all 55,309 genotyped SVs.

On the other hand, there were 122 deletions and 87 insertions that were homozygous-alternative across all samples. Out of these, 158 (75.59%) were found to overlap with the repeat regions, specifically LINE (46.10%), SINE (19.59%), LTR (15.81%), simple repeats (10.02%), and satellites (4.23%). Of note, 34,298 of 55,309 genotyped SVs overlapped with repeat regions, with proportion as follows: LINE (45.12%), SINE (28.21%), simple repeats (10.37%), LTR (2.95%), DNA transposons (1.39%), and satellites (1.39%). Here, GC contents of fixed-alternative SVs had no difference to all 55,309 genotyped SVs, p-value = 0.861 (unpaired Wilcoxon-rank test).

We excluded the two extreme SVs categories and applied an arbitrarily allele frequency (AF) threshold higher than 0.7 to identify major SV alleles. Across 14 breeds, 2330 SVs (4.21% of total genotyped SVs) were classified as major alleles. HOL, MON, and NMD had the highest numbers of breed-specific major SVs, with 951, 614, and 675, respectively (Fig. 6e), exceeding those found in other breeds such as AUB (258), BAQ (215), CHA (190), and LIM (123). Several SVs were major in most breeds but absent in specific ones, including 275 SVs missing in HOL, 252 in NMD, and 181 in MON.

Similarly, 6,648 SVs (12.01% of total genotyped SVs) were consistently rare (AF < 0.10) across all 14 breeds. HOL, MON, and NMD, had specific 1506, 873, and 1090 SVs, respectively (Fig. S18). Again, for the rare SVs, these three breeds surpassed those observed in AUB (406), BAQ (420), CHA (359), and LIM (207).

Discussion

A total of 176 samples were sequenced using LR technologies, either PacBio HiFi, ONT, or PacBio CLR, with the predominant technology being CLR. Therefore, we needed to identify the most reliable tool among CUTESV, PBSV, and SNIFFLES for consistently detecting deletions and insertions across all three sequencing platforms. Analysing a single heifer sequenced with all three technologies, we observed bias particularly for insertion detection in PacBio CLR data by SNIFFLES and CUTESV. This result was not unexpected, as the developers recommend using SNIFFLES primarily for ONT or HiFi data 39. CLR, with its longer raw reads, is known to have higher error rates, mainly due to false insertions 4. As a result, SNIFFLES detected more insertions in our CLR dataset.

For CUTESV, the F1 score derived from our heifer CLR dataset, which has approximately 42 × coverage, was slightly lower than that reported for CLR of HG002 with 69 × coverage, i.e., 0.89 for deletions and 0.83 for insertions, instead of 0.90 37. In our ONT data, PBSV slightly outperformed CUTESV and SNIFFLES. This contrasts with the results observed in human NA24385 ONT data, where CUTESV (0.90) and SNIFFLES (0.87) outperformed PBSV (0.82) 43.

This finding supports the idea that tool performance is affected by various factors such as coverage, sequencing type, SVs types, and the dataset or species being studied. In our study, PBSV demonstrated consistent performance for both deletions and insertions from HiFi, ONT and CLR reads.

We evaluated the ‘classical’ SR-based SVs detection tools of DELLY, LUMPY, and MANTA against SVs genotyping tools that also rely on SR data but incorporate a known SVs information from LR sequencing. Across deletions and insertions, the average F1 scores for these classical tools, based on an independent assessment of 148 samples, were approximately 0.50, significantly lower than genotyping tools using known SVs information from LR, such as GRAPHTYPER (0.67), PARAGRAPH (0.99), SVTYPER (0.98), and VG (0.94).

Our results also suggested VG and PARAGRAPH as the most performing tools for SV genotyping, as evidenced by their F1 scores and capability to genotype both deletions and insertions, in agreement with other findings 25. However, our average F1 scores were higher than those reported in 3 individuals of the Human Genome Structural Variation Consortium (HGSVC) with 0.72 for PARAGRAPH and 0.83 for VG 25. We further demonstrated that VG outperformed PARAGRAPH in the genotyped concordance rate independently across the 148 samples. During our analysis, we also noticed that PANGENIE 40 is able to genotype SVs from SR. However, as its algorithm is based on assembly creation from LR, we excluded it from our analysis, which focused mainly on alignment-based methods.

We incrementally added reference samples from available breeds to create SVs reference panels and subsequently re-genotyped these variants using SR from the 6 set-aside validation samples. In a human study, a leave-one-out approach was used also by generating reference variants from 10 samples while using one individual as a validation subject 40. We illustrated the impact of reference size and the number of breeds on the performance of SVs genotyping using VG. For example, the incorporation of individuals from a particular breed, (analogous to the concept of origin in humans) enhances the diversity of the reference panel and thus increases SVs recall rate for validation samples from the same breed, even if they are not directly included as the reference. This improvement occurs because SVs that exhibit significant segregation within the breed are effectively represented by other individuals of that breed.

To determine the optimal parameters for constructing the reference panel encompassing all 14 breeds, we primarily focused on the recall rate. This metric indicates the efficiency with which SVs can be recovered in six validation samples that were excluded from the panel used to construct the variation graph. We did not prioritize the precision rate because all SVs in the graph were genotyped, and for SV sites without evidence of an alternative allele, a homozygous-reference genotype was assigned using VG with the -Aa parameter. Consequently, the substantial number of SVs in the graph resulted in a decreased precision rate, though the genotyping accuracy for SVs with alternative alleles remained unaffected. Therefore, we prioritized recall rate based on the six validation samples.

We chose a simple method using JASMINE to combine SVs across multiple samples, although other tools, such as BCFTOOLS 44, SURVIVOR 2, and TRUVARI 35 could be used as well. Due to the extensive range of parameter combinations for merging SVs from reference samples using JASMINE, we restricted our analysis to two configurations: the default and adjusted parameters (see “Methods”). These were combined with eight VARCALLS thresholds to identify the optimal parameters, ultimately leading to the selection of JASMINE’s default merging parameters and a VARCALLS threshold of seven for constructing the final 14-breed panel.

We initially retained only biallelic variants from SVs discovery using PBSV on LR from individual samples, yet the merging process eventually introduced multiallelic variants and, consequently, multiallelic bubbles in the generated variation graph. These multiallelic bubbles were observed in all reference panels, from the single-breed to the 14-breeds panel. They accounted for up to 3.5% of the 53,948 SV in our final graph of 14 breeds. In contrast, the human pangenome graph shows that multiallelic bubbles account for up to 89% of the 413.809 SVs 20. The discrepancies in both SVs counts and multiallelic bubbles can be attributed to the different sample sets used: their study included 47 different human assemblies, whereas we used 14 different cattle breeds, albeit with a larger sample size of 176.

The application of variation graph on 571 SR samples enabled genotyping insertions, even with higher recall rates than deletions across 148 samples, which was difficult to detect using the classical approach from SR data 45. Thus, we expected these genotyped insertions, together with deletions, have potential implication to be discovered as many studies relied solely on the deletions for their association studies 31,32,46.

We observed two unexpected patterns in the genotype frequencies of SVs across all 571 samples, SVs fixed with homozygous-reference or homozygous-alternative across all samples. For the first pattern 635 SVs were consistently homozygous for the reference allele. These SVs exhibit relatively higher GC content than other genotyped SVs, which may have impaired SR alignment to the variation graph, resulting in insufficient evidence to call alternative alleles in any sample. In second pattern, several SVs showed a fixed allele frequency (AF) of one across all samples, suggesting that these variants may be specific to Dominette, the Hereford cow used to generate the ARS_UCD1.2 bovine reference genome.

These two extreme cases, which represent 1.5% of all genotyped SVs, cannot be solely attributed to repetitive, hard-to-genotype regions. In fact, 62% of the SVs in the final panel also overlap with repeat regions but were still successfully genotyped using SR data in the bovine reference genome.

We successfully genotyped most SVs across individuals, achieving relatively high recall rates for both deletions and insertions. However, as previously noted, SVs genotyping from SR data using variation graph by VG is not without limitations. First, VG may need more computational resources than other SV genotyping tools in different real-life scenarios. Although, we have suggested VG as the most efficient, in terms of computing time, compared to other SVs genotyping tools. This was given the condition that our analysis started from the un-aligned SR FASTQ file. However, if one considers to genotype known SVs from an already SR-aligned BAM file, possibly due to routine works in calling SNPs and InDels from WGS-SR data, then VG is sub-optimal and perhaps PARAGRAPH would be more preferred. Second, VG is dependent on an existing known SVs detected from LR sequencing. This is because known SVs are needed to build the variation graph, which in turn enables VG to genotype the corresponding SVs using SR alignments. Third, VG is not able to genotype all SVs with complete accuracy. Although recall and genotype concordance rates are relatively high, discrepancies remain in the genotyped SVs compared to those obtained directly from LR data. Our findings thus support the notion that LR sequencing is advisable for an accurate SVs detections 14,15. Nevertheless, given the higher cost and limited availability of LR data, the variation graph approach offers a practical alternative for SVs detection using SR data, as direct calling from this data type alone is expected to yield sub-optimal results.

Overall, our study presents a systematic strategy for integrating LR and SR sequencing data to comprehensively detect SVs across 14 cattle breeds. The reference panel, built from SVs identified in 176 samples, and the resulting SV panel, genotyped across 571 samples from 14 breeds, together may provide a valuable resource for population and association studies aimed at identifying genetic variants linked to important phenotypic traits. Nevertheless, there might still be a potential bias when genotyping SVs to breeds that were not represented in the 176 samples of SVs reference panel.

Material and methods

Detection of SVs from LR data

In this study, we used wholegenome LR data available for 176 animals from 14 French dairy and beef breeds (Table 1). Seven animals were sequenced with PacBio HiFi, five with Oxford Nanopore Technology (ONT), and 163 with PacBio CLR (Table S1). One sample, Charolais heifer—CHA18, was sequenced using the three LR sequencing technologies, available in European Nucleotide Archive (ENA) (see Data availability) 36. No trimming procedures were applied to the LR data output from the sequencers.

LR data were first aligned against the non-masked ARS-UCD1.2 bovine reference genome 29 using the PacBio pbmm2 47 with default parameters to generate a BAM file for each sample. Secondly, we searched for SVs using three different SV detection software, namely CUTESV v2.0.2 37, PBSV v2.6.2 38 and SNIFFLES v2.0.7 39. For both PBSV and SNIFFLES, we used default parameters. For CUTESV, we used the parameters –max_cluster_bias_INS 100 –diff_ratio_merging_INS 0.3 –max_cluster_bias_DEL 200 –diff_ratio_merging_DEL 0.5 allowing insertions to be clustered within 100 bp and deletions within 200 bp during the SV call.

Third, we filtered the SVs results in order to select only bi-allelic insertions and deletions between 50 nucleotides and 100 Kb in length, with a quality score greater than 20, well-defined SVs breakpoints, and located on the 29 autosomes and the X chromosome.

Benchmarking metrics

We used the TRUVARI benchmarking tool 35 to compare the base set, the putative true set of SVs, and the comp set, a set of SVs under evaluation. We applied the following parameters: a maximum distance between SVs of 500 bp (–refdist 500), a minimum size similarity between SVs of 0.7 (–pctsize 0.7), SV types considered during comparison (–typeignore FALSE), no reciprocal overlap (–pctovl 0), and sequence similarity ignored (–pctseq 0). This last parameter allows comparison of SVs generated by tools that do not explicitly report reference and alternative sequences in their VCF files. Our analyses are based on the summary output metrics from TRUVARI, including: recall (the proportion of SVs in the base set that are recalled in the comp set: \(TPbase/\left(TPbase+FN\right)\)), precision (the proportion of SVs in the comp set that are found in the base set: \(TPcomp/\left(TPcomp+FP\right)\)), F1 score (combining the recall and precision: \(2*\left(\left(recall*precision\right)/\left(recall+precision\right)\right)\)), and genotype concordance (the proportion of correctly genotyped SVs among all true positives, \({TPcomp}_{TPgt}/\left({TPcomp}_{TPgt}+{TPcomp}_{FPgt}\right)\)).

Evaluation of SVs detection tools on different LR data

We used BCFTOOLS v1.9 44 to extract SVs information, including length, calling support, genotype quality, and read depth, from CUTESV, PBSV and SNIFFLES based on the three LR data of CHA18. Here we used PacBio HiFi reads as the base set because of their high quality. We identified consensus SVs that were called by CUTESV, PBSV and SNIFFLES on HiFi data in two steps. First, we benchmarked the SVs called by PBSV and CUTESV to find the intersect SVs. Then, the intersect SVs were benchmarked against SVs called by SNIFFLES.

To evaluate the performance of CUTESV, PBSV, and SNIFFLES across different LR technologies, we benchmarked SVs called on HiFi, as the base set, against SVs called on ONT or CLR, as the comp set. The benchmarking was performed separately for the three callers of CUTESV, PBSV, and SNIFFLES.

For the ensemble calling of the three tools, we applied intersect and union approaches. For the former, we used the TRUVARI benchmark to obtain TP-base as the ensemble SVs. While for the union, we used JASMINE v1.1.5 41 with default parameters, jasmine –ignore_strand –mutual_distance –allow_intrasample, to merge the intended sets of SVs.

Following the developer’s recommendation to increase sensitivity and recall, we applied the –tandem-repeats parameter in the SVs discovery step of PBSV. We tested the performance on three HiFi-sequenced samples from ABO1, TAR1, and VOS1. As the input, we used a publicly available tandem repeat annotation 48 and our own generated annotation set. For the latter, tandem repeat regions were inferred using TRF v4.09.1 49 on the bovine reference genome ARS-UCD1.2 29. We applied recommended parameters set to a match weight of 2, a mismatch penalty of 7, an indel penalty of 7, a match probability of 80, an indel probability of 10, a minimum alignment score of 50, and a maximum period size of 500.

Evaluation of SR-based SVs genotyping tools

SR of CHA18 were available at 112x. SR data were also available for 154 samples (see Data availability). We down-sampled the coverage of CHA18 using SEQTK 50 to 60, 30, 15, and 10×, respectively, to mimic the coverage of other samples. SR from CHA18 and the other 154 samples, for which both LR and SR were available, were then processed subsequently to produce a BAM file for each sample. SR sequences were aligned against the ARS-UCD1.2 29 bovine reference genome with BWA MEM 51, duplications were removed and chromosomes were sorted using SAMTOOLS v1.19 44, and base quality score recalibration was performed using GATK v4.5 52 based on the known ~ 106 million SNPs and InDels from the 1000 Bulls genome project 53. The BAM file was then processed by DELLY v1.2.6 10, LUMPY 11 via SMOOVE v0.2.8 54, and MANTA 9 to call SVs. We only retained deletions called by LUMPY and DELLY with the additional PASS flag for DELLY. We kept both deletions and insertions produced by MANTA.

For each sample, the SVs detected by PBSV on the CLR data were used as the base set. This base set was also used as known sites to genotype SVs from SR data using GRAPHTYPER v2.7.7 21, PARAGRAPH v2.3 22, SVTYPER v0.7.1 23 and VG v1.56 24. The first three tools take a BAM file as input, the same as used by LUMPY, DELLY and MANTA. With default parameters, GRAPHTYPER generates numerous chunks of genotyped SVs across the genome. We used the BCFTOOLS –naive concat function to combine all the chunks, sort them and retain only biallelic SVs defined by the “AGGREGATED” SV model with a size greater than 50 bp and a maximum alternative allele support of 0.5. For PARAGRAPH, we used information on the average SR length and alignment coverage from the final BAM file. For SVTYPER, we added information from CIPOS and CIEND with values of -100,100; and CIPOS95 and CIPOS95 END with 0,0 to the vcf info field of known SV sites. We genotyped SVs using svtyper-sso with the parameters –max_reads 1,000,000 –max_ci_dist 4 –split_weight 1 –disc_weight 1. We then removed all SVs with missing genotypes defined by PARAGRAPH and SVTYPER.

For SVs genotyping using VG in each sample, we first constructed the variation graph for all chromosomes separately based on the individual’s base set. These separate chromosome graphs were then merged into a whole-genome graph by re-setting their node ids to follow the order of the 29 autosomes plus the X chromosome. The whole-genome graph was then indexed to an .XG, converted to .GFA, and graph libraries (.GBZ, .MIN, and .DIST) were generated to facilitate the alignment. SR were mapped against the graph using VG GIRAFFE 55 to create a GAM file. The alignment and graph were chunked by chromosome. SNARLS file, which contains breakpoint information of SVs embedded in the graph, and PACK file, which contains the post-alignment coverage index with a minimum mapping quality of 5 (–min-mapq 5), were created from the chunked graph and alignment, respectively. SVs genotyping was then performed based on the evidence in these SNARLS and PACK files using vg call sample_chr.vg -k sample chr.pack -r sample chr.snarls -s sample.id. We retained only biallelic genotyped SVs across autosomes and X chromosome. By default, VG only outputs SVs with heterozygous or homozygous alternative genotypes. As the output does not include SVTYPE information, we classify genotyped SVs as deletions if the REF allele is longer than 50 bp and the ALT allele is one bp, or as insertions if the REF allele is one bp and the ALT allele is longer than 50 bp.

We performed benchmarking on the genotyped SVs recovered from SR by seven tools, using them as comp sets against the base set for each individual. The benchmarking was carried out separately for deletions and insertions.

Extending the number and diversity of reference SV panels

We used JASMINE v1.1.5 41 to merge the base sets of reference samples with the default parameters jasmine –ignore_strand –mutual_distance –allow_intrasample –output_genotypes for each panel. For each panel, only SVs with a minimum number of alternative alleles of four were retained using VCFTOOLS 56. The SRs of the validation samples were aligned to the graphs and the genotyped SVs were called. The genotyped SVs, here as comp sets, were then compared against the base set of each validation individual. We used the deconstruct option of VG for decomposing the graph referring to positions in the ARS-UCD1.2 and counted the bubbles generated within.

Optimum parameters for building a reference panel of SVs across 14 breeds

We kept the 6 validation and 145 reference samples from the previous section. Apart from the default JASMINE parameters, we also modified the merging parameters with jasmine –ignore_strand –mutual_distance –allow_intrasample –max_dist_linear = 0.1 –min_dist = 50. In short, compared to the default parameters, JASMINE only merged SVs within a distance proportional to 0.1 of their size, instead of 0.5, and with a minimum distance of 50, instead of 100. These parameters were recommended by the JASMINE developer in a github issue and increased the stringency of SV merging.

In addition to the minimum alternative-allele counts of four, we also applied eight thresholds of 2, 3, 5, 7, 13, 20, 27, and 36 from VARCALLS after SVs merging. This information is output by JASMINE and indicates the number of variant calls supporting the merged SVs. Thus, 8 × 2 reference panels were created from these adjustments, on which variation graphs were constructed. Subsequently, the SR data of the validation samples were aligned to these variation graphs and subjected to VG genotyping. When applying VARCALLS thresholds, we removed SVs from the reference panel. Subsequently, we adjusted the base set of validation samples by also excluding the SVs that were removed due to VARCALLS filtering. We benchmarked the genotyped SVs (as comp sets) against the adjusted base set of validation samples using TRUVARI.

Genotyping SVs in a large cohort of SR samples

We created the final reference panel using the SV sets called by PBSV on LR of 176 samples from 14 breeds (Table 1; Table S1). We merged the SVs with JASMINE using default parameters with a VARCALLS threshold of seven, followed by the creation of the final variation graph. SR from 571 samples were aligned against the graph and then processed for SVs genotyping. Collection of these SR samples has been described elsewhere 57 (see Data availability). We used the –Aa parameter in the VG call to output all genotypes, including missing or homozygous reference, for all SVs in the final panel using vg call panel chr.vg -Aa -k sample chr.pack -r sample chr.snarls -s sample.id. Annotation of the genotyped SVs was carried out using snpEff 42 based on the bovine annotation of ARS-UCD1.2 v99.

PCA was performed on the genotyped SVs using PLINK v1.9 58 with 20 eigenvectors. For 148 samples with LR and SR available, we benchmarked the genotyped SVs, as the comp set, against the base set of each sample, respectively. We used VCFTOOLS 56 to infer the frequency of genotyped SVs. We used UpSet 59 to characterize major (AF > 0.7) or rare (AF < 0.1) SVs in all samples. We used BEDTOOLS 60 to intersect SVs regions of interest to repeat annotations 48 of the bovine reference genome. R 61 and GGPLOT2 62 were used mainly for data handling and plotting. The pipelines used in the analysis have also been made publicly available (see Data availability).