Chromosome-length genome assembly and structural variations of the primal Basenji dog (Canis lupus familiaris) genome

Richard J Edwards; Matt A Field; James M Ferguson; Olga Dudchenko; Jens Keilwagen; Benjamin D Rosen; Gary S Johnson; Edward S Rice; La Deanna Hillier; Jillian M Hammond; Samuel G Towarnicki; Arina Omer; Ruqayya Khan; Ksenia Skvortsova; Ozren Bogdanovic; Robert A Zammit; Erez Lieberman Aiden; Wesley C Warren; J William O Ballard

doi:10.1186/s12864-021-07493-6

. 2021 Mar 16;22:188. doi: 10.1186/s12864-021-07493-6

Chromosome-length genome assembly and structural variations of the primal Basenji dog (Canis lupus familiaris) genome

Richard J Edwards ^1,^#, Matt A Field ^2,^3,^#, James M Ferguson ^4,^#, Olga Dudchenko ^5,^6,^7,^#, Jens Keilwagen ⁸, Benjamin D Rosen ⁹, Gary S Johnson ¹⁰, Edward S Rice ¹¹, La Deanna Hillier ¹², Jillian M Hammond ⁴, Samuel G Towarnicki ¹, Arina Omer ^5,⁶, Ruqayya Khan ^5,⁶, Ksenia Skvortsova ^13,¹⁴, Ozren Bogdanovic ^1,¹³, Robert A Zammit ¹⁵, Erez Lieberman Aiden ^5,^6,^7,^16,^17,^✉, Wesley C Warren ^18,^✉, J William O Ballard ^19,^20,^✉

¹School of Biotechnology and Biomolecular Sciences, University of New South Wales, Sydney, NSW 2052 Australia

²Centre for Tropical Bioinformatics and Molecular Biology, Australian Institute of Tropical Health and Medicine, James Cook University, Cairns, QLD 4878 Australia

³John Curtin School of Medical Research, Australian National University, Canberra, ACT 2600 Australia

⁴Kinghorn Center for Clinical Genomics, Garvan Institute of Medical Research, Victoria Street, Darlinghurst, NSW 2010 Australia

⁵The Center for Genome Architecture, Department of Molecular and Human Genetics, Baylor College of Medicine, Houston, TX USA

⁶Department of Computer Science, Rice University, Houston, TX USA

⁷Center for Theoretical and Biological Physics, Rice University, Houston, TX USA

⁸Julius Kühn-Institut, Erwin-Baur-Str, 27 06484 Quedlinburg, Germany

⁹Animal Genomics and Improvement Laboratory, Agricultural Research Service USDA, Beltsville, MD 20705 USA

¹⁰Department of Veterinary Pathobiology, University of Missouri, Columbia, MO 65211 USA

¹¹Department of Surgery, University of Missouri, Columbia, MO 65211 USA

¹²Genome Sciences, University of Washington, Seattle, WA 98195 USA

¹³Genomics and Epigenetics Division, Garvan Institute of Medical Research, Victoria Street, Darlinghurst, NSW 2010 Australia

¹⁴St Vincent’s Clinical School, Faculty of Medicine, University of New South Wales, Sydney, NSW 2010 Australia

¹⁵Vineyard Veterinary Hospital, 703 Windsor Rd, Vineyard, NSW 2765 Australia

¹⁶Faculty of Science, UWA School of Agriculture and Environment, University of Western Australia, Perth, WA 6009 Australia

¹⁷Shanghai Institute for Advanced Immunochemical Studies, ShanghaiTech University, Shanghai, China

¹⁸Department of Animal Sciences, University of Missouri, Columbia, MO 65211 Australia

¹⁹Department of Ecology, Environment and Evolution, La Trobe University, Melbourne, Victoria 3086 Australia

²⁰School of Biosciences, University of Melbourne, Parkville, Victoria 3052 Australia

^✉

Corresponding author.

Contributed equally.

PMCID: PMC7962210 PMID: 33726677

Abstract

Background

Basenjis are considered an ancient dog breed of central African origins that still live and hunt with tribesmen in the African Congo. Nicknamed the barkless dog, Basenjis possess unique phylogeny, geographical origins and traits, making their genome structure of great interest. The increasing number of available canid reference genomes allows us to examine the impact the choice of reference genome makes with regard to reference genome quality and breed relatedness.

Results

Here, we report two high quality de novo Basenji genome assemblies: a female, China (CanFam_Bas), and a male, Wags. We conduct pairwise comparisons and report structural variations between assembled genomes of three dog breeds: Basenji (CanFam_Bas), Boxer (CanFam3.1) and German Shepherd Dog (GSD) (CanFam_GSD). CanFam_Bas is superior to CanFam3.1 in terms of genome contiguity and comparable overall to the high quality CanFam_GSD assembly. By aligning short read data from 58 representative dog breeds to three reference genomes, we demonstrate how the choice of reference genome significantly impacts both read mapping and variant detection.

Conclusions

The growing number of high-quality canid reference genomes means the choice of reference genome is an increasingly critical decision in subsequent canid variant analyses. The basal position of the Basenji makes it suitable for variant analysis for targeted applications of specific dog breeds. However, we believe more comprehensive analyses across the entire family of canids is more suited to a pangenome approach. Collectively this work highlights the importance the choice of reference genome makes in all variation studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12864-021-07493-6.

Keywords: Canine genome, Domestication, Comparative genomics, Artificial selection

Background

Dogs were the first animals to be domesticated by humans some 30,000 years ago [1] and exhibit exceptional levels of breed variation as a result of extensive artificial trait selection [2]. It is not clear whether dogs were domesticated once or several times, though the weight of accumulating evidence suggests multiple events [3–9]. By establishing genome resources for more ancient breeds of dog, we can explore genetic adaptations perhaps unique to the modern dog breeds. Basenjis are an ancient breed that sits at the base of the currently accepted dog phylogeny [10]. Basenji-like dogs are depicted in drawings and models dating back to the Twelfth Dynasty of Egypt [11] and they share many unique traits with pariah dog types. Like dingoes and New Guinea Singing dogs (NGSD), Basenjis come into oestrus annually—as compared to most other dog breeds, which have two or more breeding seasons every year. Basenjis, dingoes and NGSDs are prone to howls, yodels, and other vocalizations over the characteristic bark of modern dog breeds. One explanation for the unusual vocalisation of the Basenji is that the larynx is flattened [12]. The shape of the dingo and NGSD larynx is not reported.

Basenjis were originally indigenous to central Africa, wherever there was tropical forest. Primarily, what is now the DRC Congo, Southern Sudan, Central African Republic and the small countries on the central Atlantic coast. Today their territory has shrunk to the more remote parts of central Africa. The Basenji probably made its debut in the western world in around 1843. In a painting of three dogs belonging to Queen Victoria and Prince Albert entitled “Esquimaux, Niger and Neptune”, Niger is clearly a Basenji. In total, 71 Basenjis have been exported from Africa and, to date, ~ 56 have been incorporated into the registered Basenji breeding population.

The first dog genome to be sequenced was of Tasha the Boxer [13], which was a tremendous advance and continues to be the resource guiding canine genomics research today. The Boxer is a highly derived brachycephalic breed that has been subjected to generations of artificial selection. Further, due to its discontiguous sequence representation it has been difficult to accurately detect structural variations (SVs) in other domestic dog breeds. Now, a new generation of breed-specific chromosome-level genome reference assemblies are becoming available (5 breeds in October 2020 according to the NCBI assembly archive). For example, we previously published a chromosome-level German Shepherd dog (GSD) genome assembly (CanFam_GSD) that is comprised of only 410 scaffolds and 716 contigs [14].

Here, we first report the sequence and de novo assembly of two Basenji genomes, female and male. We then compare these assemblies with the Boxer (CanFam3.1) [15] and GSD (CanFam_GSD) [14]. We conduct pairwise comparisons and report single-nucleotide variants (SNVs) and SVs between Basenji, Boxer and GSD. We distinguish an SNV as a variation in a single nucleotide without any limitations on its frequency. SV comprises a major portion of genetic diversity and its biological impact is highly variable. Chen et al. [16] used high-resolution array comparative genome hybridization to create the first map of DNA copy number variation (CNV) in dogs. Many canine CNVs were shown to effect protein coding genes, including disease and candidate disease genes, and are thus likely to be functional. In this study, we find all types of genetic variation are impacted by the choice of reference genome. The basal position of the Basenji makes it useful as a general reference for variant analysis, but the generation of clade-specific genomes is likely to be important for canine nutrition and disease studies. We recommend a pan-genome approach for comprehensive analyses of canid variation.

Results

Basenji female assembly, CanFam_Bas

The female Basenji, China (Fig. 1a), was initially assembled from 84.5 Gb Oxford Nanopore Technologies (ONT) PromethION reads (approx. 35x depth based on a 2.41 Gb genome size) using Flye (v2.6) [17, 18] and subjected to long read polishing with Racon v1.3.3 [19] and Medaka 1.03 [20] (Supplementary Fig. 1A, Additional File 1). Additional short read Pilon [21] error-correction was performed with 115.1 Gb (approx. 47.7x) BGIseq data. Hi-C proximity ligation was used with the DNA zoo pipeline [22–24] to scaffold 1657 contigs into 1456 scaffolds, increasing the N50 from 26.3 Mb to 63.1 Mb and decreasing the L50 from 33 to 14 (Figs. 2 and 3, Supplementary Table 1, Additional File 2). Scaffolds were gap-filled by PBJelly (pbsuite v.15.8.24) [25] using the ONT data, reducing the number of gaps from 348 to 148 and the number of scaffolds to 1407. Following a final round of Pilon [21] BGIseq-polishing, scaffolds were mapped onto the CanFam3.1 [13] using PAFScaff v0.40 [14, 26]. Diploidocus v0.9.0 vector filtering [27] removed one 5.7 kb contig and masked a 3.3 kb region of Chromosome X as lambda phage (J02459.1) contamination. Seven rounds of iterative Diploidocus tidying of the remaining sequences removed 277 (832 kb) as low coverage/quality and 481 (1.58 Mb) as probable haplotigs, retaining 483 core scaffolds and 165 probable repeat-heavy sequences [14] as China v1.0 (Fig. 3, Supplementary Fig. 2, Additional File 1, Supplementary Table 2, Additional File 2).

Fig. 2 — Contact matrices generated by aligning the CanFam_Bas (China) Hi-C data set to the genome assembly a. before the Hi-C upgrade (draft assembly). b. After Hi-C scaffolding (End-to-end assembly)

Fig. 3 — Key contiguity, quality and completeness metrics for different assembly stages and comparison dog genomes. Square, pre-scaffolded China; Diamond, scaffolded China; Triangle, complete assembly; Circle, main chromosome scaffolds only; Blue, China; Purple, Wags; Red, CanFam_GSD; Green, CanFam3.1. a. Contig (open) and scaffold (filled) numbers. b. Contig (open) and scaffold (filled) N50. c. Contig (open) and scaffold (filled) L50. d. Genome completeness estimated by BUSCO v3 (filled) and Merqury (open). e. The percentage of missing BUSCO genes (filled) and BUSCOMP genes (those found to be Complete in any assembly) (open). f. Schematic of China assembly workflow. CanFam_Bas is China v1.2

Genome assembly correction

Two pairs of fused chromosomes in China v1.0 were incorrectly joined by PBJelly. Pre-gap-filled HiC scaffolds were mapped onto the assembly using Minimap2 v2.17 [28] and parsed with GABLAM v2.30.5 [29] to identify the gap corresponding to the fusion regions. These were manually separated into four individual chromosomes, gap lengths standardised, and scaffolds re-mapped onto CanFam3.1 using PAFScaff v0.4.0. D-GENIES [30] analysis against CanFam_GSD chromosomes confirmed that PBJelly had incorrectly fused two pairs of chromosomes: chromosomes 8 with 13, and chromosome 18 with 30. These were manually separated and the assembly re-mapped onto CanFam3.1 as China v1.1. PAFScaff assigned 112 scaffolds to chromosomes, including 39 nuclear chromosome-length scaffolds.

It was observed that the mitochondrial chromosome was missing and China v1.1 Chromosome 29 contained a 33.2 kb region consisting of almost two complete copies of the mitochondrial genome that were not found in other dog genome assemblies. The 26 ONT reads that mapped onto both flanking regions were reassembled with Flye v.2.7.1 [17, 18] into a 77.2 kb chromosome fragment, which was polished with Racon v1.3.3 [19] and Medaka 1.03 [20]. This was mapped back on to the Chromosome 29 scaffold with GABLAM v2.30.5 [29] (blast+ v2.9.0 megablast [31, 32]) and the mitochondrial artefact replaced with the corrected nuclear mitochondrial DNA (NUMT) sequence. Finally, scaffolds under 1 kb were removed to produce the China v1.2 nuclear genome that we name CanFam_Bas.

Mitochondrial genome assembly

In total, 4740 ONT reads (52.1 Mb) mapping on to mtDNA were extracted. To avoid NUMT contaminants, a subset of 80 reads (1.32 Mb) with 99% + mtDNA assignment and 99% + mtDNA coverage, ranging in size from 16,197 kb to 17,567 kb, were assembled with Flye 2.7b-b1526 [17, 18] (genome size 16.7 kb) into a 33,349 bp circular contig consisting of two mtDNA copies. This contig was polished with Racon [19] and Medaka [20], before being re-circularised to a single-copy matching the CanFam3.1 mtDNA start position. After final Pilon [21] correction of SNPs and indels, the 16,761 bp mitochondrial genome was added to the CanFam_Bas assembly.

CanFam_Bas (China) reference genome

The resulting chromosome-length CanFam_Bas reference genome is 2,345,002,994 bp on 632 scaffolds with 149 gaps (76,431 bp gaps) (Table 1). The 39 nuclear plus mitochondrial chromosome scaffolds account for 99.3% of the assembly and show a high level of synteny with CanFam3.1 and CanFam_GSD (Fig. 4). CanFam_Bas represents the most contiguous dog chromosomes to date, with a contig N50 of 37.8 Mb and contig L50 of 23, which is slight improvement over CanFam_GSD and considerably more contiguous than the standard dog reference genome, CanFam3.1 (Fig. 3, Table 1). The completeness and accuracy of the genome as measured by BUSCO v3 [33] (laurasiatherian, n = 6253) is also superior to CanFam3.1 and approaches that of CanFam_GSD (92.9% Complete, 3.75% Fragmented, 3.34% Missing).

Table 1.

Genome assembly and annotation statistics for Basenji assemblies vs CanFam3.1 and CanFam_GSD

	CanFam_Bas (China)	Wags	CanFam3.1	CanFam_GSD
Total sequence length	2,345,002,994	2,410,429,933	2,410,976,875	2,407,308,279
Total ungapped length	2,344,926,563	2,410,291,233	2,392,715,236	2,401,163,822
Number of scaffolds	632	2243	3310	430
Scaffold N50	64,291,023	61,087,166	45,876,610	64,346,267
Scaffold L50	14	16	20	15
Number of contigs	780	3630	27,106	736
Contig N50	37,759,230	3,131,423	267,478	20,914,347
Contig L50	23	217	2436	37
No. chromosomes	40	40	40	40
Percentage genome in main chromosomes	99.3%	94.8%	98.3%	96.5%
BUSCO complete (genome)	92.9% (1.14% Duplicated)	91.5% (1.31% Duplicated)	92.2% (1.17% Duplicated)	93.0% (1.38% Duplicated)
BUSCO fragmented (genome)	3.74%	4.53%	4.03%	3.73%
BUSCO missing (genome)	3.34%	3.98%	3.73%	3.37%
BUSCO complete (proteome)	98.5% (1.9% Duplicated)	97.8% (2.4% Duplicated)	95.1% (1.0% Duplicated)	98.9% (2.4% Duplicated)
BUSCO fragmented (proteome)	1.2%	1.5%	1.9%	1.0%
BUSCO missing (proteome)	0.3%	0.7%	3.0%	0.1%

Open in a new tab

Fig. 4 — D-GENIES synteny plots of main chromosome scaffolds for three dog genome assemblies against CanFam_Bas. In each case, CanFam_Bas (China v1.2) is on the x-axis and the comparison assembly on the y-axis. Gridlines demarcate scaffolds. Thick black lines indicate regions of genomic alignment. a. All-by-all main chromosome scaffold alignments with (i) CanFam_GSD, (ii) CanFam_3.1, and (iii) Wags. b. Main chromosome 9 scaffold alignment with (i) CanFam_GSD, (ii) CanFam_3.1, and (iii) Wags

Methylomic identification of putative regulatory elements

Additionally, we profiled whole genome methylation of Basenji’s blood DNA using MethylC-seq [34]. Numbers of unmethylated and highly methylated CpG sites in Basenji’s genome were similar to that of GSD (Supplementary Fig. 3A, Additional File 1). Importantly, high resolution DNA methylation data can be utilised to identify the putative regulatory elements in a given tissue type. That is, CpG-rich unmethylated regions (UMRs) mostly correspond to gene promoters, while CpG-poor low-methylated regions (LMRs) correspond to distal regulatory elements, both characterised by DNAse I hypersensitivity [35]. Using MethylSeekR algorithm [36], we performed the segmentation of Basenji DNA methylome and discovered 20,660 UMRs and 54,807 LMRs (Supplementary Fig. 3B,C, Additional File 1), in line with previously reported numbers in mammalian genomes [14, 36, 37]. Genome-wide and locus-specific CpG methylation called by MethylC-seq correlated strongly with that called directly from the ONT data (Supplementary Fig. 3D-F, Additional File 1), confirming the robustness of determined DNA methylation profile of the blood DNA.

Male basenji assembly, wags

For the male Basenji, Wags, (Fig. 1b) we generated Pacific Bioscience Single Molecule Real Time (SMRT) sequences to approximately 45x genome coverage and assembled the genome to ungapped size of 2.41 Gb (Supplementary Fig. 1B, Additional File 1). Assembly contiguity metrics of 3630 total contigs show N50 contig and scaffold lengths of 3.1 and 61 Mb length, respectively (Table 1). Wags alignment to China revealed a high level of synteny. However, the Wags assembly of the X chromosome is smaller in size (59 Mb vs 125 Mb) and shows multiple rearrangements as a result of lower sequence coverage on the sex chromosomes (~21x). We were unable to accurately place 124.4 Mb of Wags sequence on 2204 scaffolds (2210 contigs), including 651 contigs with a total length of 45.6 Mb mapped on to the CanFam3.1 X chromosome by PAFScaff. Therefore, all comparative analyses reported herein were done with CanFam_Bas. In addition, the Wags assembly includes 3.6 Mb of the Basenji dog Y for future comparative studies of this unique chromosome.

Genome annotation

The CanFam_Bas and Wags assemblies were annotated using the homology-based gene prediction program GeMoMa v1.6.2beta [38] and nine reference species [14]. In total, CanFam_Bas and Wags had similar numbers of predicted protein-coding genes at 27,129 (68,251 transcripts) and 27,783 (65,511) transcripts, respectively (Supplementary Table 3, Additional File 2). Analysing the longest protein isoform per gene with BUSCO v3 [33] (laurasiatherian, n = 6253, proteins mode), CanFam_Bas was measured to be 98.5% complete (1.9% duplicated) and Wags was measured as 97.8% complete (2.4% duplicated). Both proteomes compare favourably with CanFam3.1 in terms of completeness (Table 1). To correct for differences introduced by the annotation method, CanFam3.1 was annotated with the same GeMoMa pipeline. Approximately 90% of the Quest For Orthologues (QFO) reference dog proteome [39] is covered by each GeMoMa proteome, confirming comparable levels of completeness (Supplementary Table 3, Additional File 2).

When the CanFam_Bas GeMoMa proteome was compared to Wags, CanFam3.1 and CanFam_GSD, over 91% genes had reciprocal best hits for at least one protein isoform (Supplementary Table 3, Additional File 2). To investigate this further, the Wags, CanFam3.1 and CanFam_GSD genomes were mapped onto CanFam_Bas and the coverage for each gene calculated with Diploidocus v0.10.0. Of the 27,129 predicted genes, 26,354 (97.1%) are found at least 50% covered in all four dogs, whilst only 30 (0.11%) are completely unique to CanFam_Bas. In total, Wags is missing 302 predicted genes, CanFam_GSD is completely missing 95 predicted genes, and CanFam3.1 is missing 211 predicted genes (Table 2). A considerably greater proportion of the missing genes in Wags (64.2% versus 11.4% in CanFam3.1 and 15.8% in CanFam_GSD) were on the X chromosome. To test for artefacts due to assembly errors we mapped the long read data for Wags and CanFam_GSD onto CanFam_Bas. Only 7 of the 302 missing Wags genes (2.3%) had no long read coverage, whilst 21/95 (22.1%) of genes missing in CanFam_GSD were confirmed by an absence of mapped long reads.

Table 2.

Predicted copy numbers for CanFam_Bas GeMoMa genes based on A. assembly mapping, and B. long read mapping

A. Dog	Missing	Partial (< 50%)	Single (1n)		Duplicate (2n)	3n+
CanFam_Bas	0 (0)	0 (0)	27,129 (25788)		0 (0)	0 (0)
CanFam_Wags	302 (108)	120 (58)	26,103 (25035)		486 (472)	118 (115)
CanFam3.1	211 (187)	167 (161)	26,404 (25125)		251 (223)	96 (92)
CanFam_GSD	95 (80)	48 (42)	26,586 (25304)		299 (266)	101 (96)
B. Data	0n	0.5n	1n	1.5n	2n	2.5n+
CanFam_Bas (ONT)	2 (2)	1257 (1201)	24,116 (22954)	1508 (1403)	80 (68)	166 (160)
CanFam_Wags (PacBio)	7 (2)	4717 (3476)	21,140 (21049)	1028 (1024)	79 (79)	158 (158)
CanFam_GSD (ONT)	21 (18)	1893 (1795)	22,412 (21350)	2503 (2349)	109 (98)	191 (178)

Open in a new tab

Figures in brackets exclude predicted genes on X chromosome

Amylase copy number

Two copies of the Amy2B gene were identified in a tandem repeat on Chromosome 6 of the CanFam_Bas assembly. The single-copy read depth for CanFam_Bas, calculated as the modal read depth across single copy complete genes identified by BUSCO v3 [33], was estimated to be 34x. This was verified by BUSCO complete genes, which gave mean predicted copy numbers of 1.008 ± 0.005 (95% C.I.) (Supplementary Fig. 4, Additional File 1). The two complete Amy2B coding sequence copies had a mean depth of 97.5x, equating to 2.87 N, or a total copy number estimate of 5.78 N (2 × 97.5 / 34). The full CanFam_GSD Amy2B repeat region was also found in two complete copies with a mean depth of 98.1x, estimating 5.77 copies (2 × 98.1 / 34). Similar results were obtained restricting analysis to reads at least 5 kb (6.01 gene copies; 5.98 region copies) or 10 kb (6.18 gene copies; 6.04 region copies) to minimise repeat-based fluctuations in read depth. In contrast, droplet digital PCR (ddPCR) estimated that the Basenji China had 4.5 copies per individual chromosome. This slight difference suggests that the raw sequence data slightly overestimated copies or the ddPCR primers did not capture all the genetic variation.

Wags assembly has a single copy of the Amy2B region, which includes 90% of the Amy2B coding sequence (data not shown). Single-copy depth analysis of Wags estimates 4.97 (90% at 253.8x) and 5.51 (100% at 253.5x) copies of the AMY2B coding sequence and tandem repeat unit, respectively. To estimate copy number in other Basenji dogs, short read data was downloaded from SRA for 11 Basenjis (Supplementary Table 4, Additional File 2) and mapped onto CanFam_BAS. AMY2B copy number was estimated as the ratio of mean read depth per AMY2B gene versus the mean depth for the whole genome. Estimated copy numbers ranged from 3.93 to 6.79, with a mean of 5.25 (Supplementary Table 4, Additional File 2). The same analysis was performed on China BGI data, yielding an estimate of 4.38 copies, consistent with the ddPCR results.

Nuclear mitochondrial DNA fragments

During the assembly of the female Basenji genome (China v1.0), the mitochondrial genome was erroneously assembled into a NUMT fragment on chromosome 29. A blastn search (e < 10–4 [40]) identified 291 putative NUMT fragments, ranging in size from 34 bp to 6580 bp, forming 212 NUMT blocks from 34 bp to 16,577 bp (Supplementary Table 5,6, Additional File 2). Fragments total 190.5 kb (approx. 11.4 mtDNA copies) and span the entire mtDNA with reasonably even coverage and no obvious bias (Supplementary Fig. 5A, Additional File 1), except for low coverage in a region of the D-loop as has been previously reported in primates [41]. All 291 NUMT fragments are well-supported with at least 3 reads that span the entire NUMT plus at least 5 kb each side (Supplementary Table 5, Additional File 2). Only 1 NUMT was not found to be full-length in CanFam_GSD (Supplementary Fig. 5B, Additional File 1). An additional 26 NUMTs are partially covered in CanFam3.1 and 9 are entirely absent. Whilst this could represent a breed difference, 19 of the 35 additional incomplete NUMTs in CanFam3.1 (65.5%) are also incomplete in Wags, whilst Wags has a further ten incomplete NUMTs that are present in CanFam3.1 (Supplementary Table 5, Additional File 2). This is consistent with these regions being generally harder to assemble and/or align. Further analyses of these regions may provide insight into domestic dog genealogies.

Whole genome assembly comparisons

To discover unique large-scale structural differences in assembled genomes of the three breeds – Basenji, German Shepard and Boxer – we performed pairwise alignments of CanFam_Bas, CanFam3.1 and CanFam_GSD. Overall, genome synteny was maintained and there were limited large scale genomic rearrangements observed (Fig. 4a). There was, however, a large inversion in CanFam3.1 that was not present in CanFam_Bas or CanFam_GSD (Fig. 4b). To investigate this further, we aligned CanFam_Bas against Wags (Fig. 4a iii). As expected, there was no inversion on Chromosome 9 (Fig. 4b iii), suggesting the inversion or perhaps an assembly error, occurs in CanFam3.1.

Long read structural variant detection

To generate a conservative set of high-quality structural variants, the consensus of ONT and SMRT long read sequences data from both Basenji (China female ONT and Wags male SMRT) and GSD (Nala female ONT and SMRT [14]) was mapped onto Basenji (CanFam_BAS), Boxer (CanFam3.1) and GSD (CanFam_GSD) reference genomes (Supplementary Table 7, Additional File 2). One difference in the data sets is that both ONT and SMRT reads from GSD were sourced from the same individual dog whereas for Basenji the ONT and SMRT reads are from different individual Basenji dogs. These high-quality SVs were additionally annotated for their overlap to both genes and exons. Of these high-quality ONT/SMRT consensus SVs, the Basenji long reads overlap 814 CanFam3.1 exons and 568 CanFam_GSD exons while the GSD long reads overlap 825 CanFam3.1 exons and 495 CanFam_Bas exons. In total, these SVs represent 92.19 Mb, 97.21 Mb, and 78.99 Mb of deleted sequence and 4.11 Mb, 4.09 Mb, and 7.69 Mb of inserted sequence for the CanFam_Bas, CanFam3.1, and CanFam_GSD assemblies respectively.

To reduce the impact of small indels arising from local mis-assembly errors, the high-quality consensus SVs were further reduced to those over 100 bp in length (Fig. 5). Overall, we observe similar number of total SV calls from the Basenji long reads relative to CanFam_GSD and the GSD long reads relative to CanFam_BAS. Both breeds had a substantially larger number of consensus SV calls against the CanFam3.1 reference, with Basenji long reads generating more SVs calls than GSD long reads. We next overlapped the CanFam SV calls relative to CanFam3.1 and found 18,063 long read deletion calls overlapped between Basenji and GSD. For Basenji this represented 70.00% of the total 25,260 Basenji deletions while for GSD this represented 73.25% of the total 24,138 GSD deletions (Fig. 5b). Insertions were fewer in number and degree of overlap, however we still found 5146 overlapping insertions between Basenji and GSD long reads representing 33.33% of the total 15,434 Basenji insertions and 36.46% of the total 14,111 GSD insertions (Fig. 5c). The high degree of overlap in SVs from GSD and Basenji relative to CanFam3.1 represent Boxer-specific SVs or potential issues with the current canid reference assembly.

Fig. 5 — Consensus structural variant calls for combined ONT and PacBio data. High-quality set of consensus structural variant (SV) calls generated from the intersection of the ONT and PacBio SV calls for each breed versus reference comparison, limited to SVs > 100 bp long. a. Total numbers of SVs called from Basenji CanFam_Bas reads (red) versus CanFam3.1 and CanFam_GSD, and German Shepherd long reads (green) versus CanFam3.1 and CanFam_Bas. b. Numbers and overlap of consensus deletion calls for Basenji reads (blue) and GSD reads (green) versus CanFam3.1. c. Numbers and overlap of consensus insertion calls for Basenji reads (blue) and GSD reads (green) versus CanFam3.1

Short read mapping, SNV / small indel detection

Mapping multiple individuals to a reference genome for variant calling is standard practice in population genomics and is known to be prone to biases in the reference genome. To investigate whether differences identified in variant analyses were due to the Basenji being a basal breed or due to assembly quality difference, short read data from 58 different dog breeds belonging to sixteen different well-supported clades from Parker et al. [10] (Supplementary Table 8, Additional File 2) were mapped on to three reference genomes Basenji (CanFam_BAS), Boxer (CanFam3.1) and GSD (CanFam_GSD). Large-scale structural differences between breeds would be expected to significantly affect read mapping efficiencies for closely-related versus distantly-related breeds, whilst missing assembly data would be expected to result in a systematic reduction in mapping across all breeds. In our analysis, we observe such systematic and breed specific changes in both the number of mapped reads and variants detected.

For the systematic changes, overall trends are exhibited in the total percentage of reads mapped across the three references, with the highest percent of reads mapping to CanFam_GSD, followed closely by CanFam_BAS and the lowest percent of reads mapping to CanFam3.1 (Fig. 6a). ANOVA shows these differences are significant (F_{2, 171} = 819.53, P < 0.0001). To investigate this result further and test for interactions we focused upon breeds within each of the monophyletic clades close to or associated with the three reference genomes. Specifically, we included the short-read sequences from the four breeds in the Asian Spitz clade as this close to the basal Basenji lineage. We also included the six breeds European Mastif clade containing the Boxer and three breeds within the New World clade containing the GSD. Overall, the CanFam_Bas and CanFam_GSD performed equally well while the relative mapping was lowest for CanFam3.1 (Fig. 6b). Once again ANOVA shows this result is significant (F_{8, 30} = 32.01, P < 0.0001). Next, we considered the capacity of each reference genome to detect SNV’s and indels. In this case CanFam_BAS detected higher number of changes than did either CanFam3.1 or CanFam_GSD (Fig. 6c and d; ANOVA F_{2, 171} = 30.71, and F_{2, 171} = 12.08, respectively with P < 0.0001 for each). In combination these data attest to the quality of the CanFam_Bas assembly. There is a noticeable depletion of variant calls for the reference breed (Supplementary Fig. 6, Additional File 1), but there were no significant interactions between clades and reference genome (data not shown).

Fig. 6 — Comparative short read mapping and single nucleotide variant calling for 58 dog breeds versus three reference genomes: CanFam_Bas (Bas), CanFam3.1 (3.1) and CanFam_GSD (GSD). a. Relative percentage of reads mapped to each reference calculated by subtracting the mean percentage of reads for the three reference genomes from the number of reads for each reference. b. Relative percentage of reads mapped to each reference for three closely related clades determined from Parker et al. [10]. The Basenji is closely related to the Asian Spitz clade. The Boxer is a member of the European Mastiff clade and the GSD is a member of the New world clade. c. The number of SNVs. d. The number of small indels

Discussion

In this manuscript we present a reference-quality assembly of a female Basenji (China), which we designate CanFam_Bas. We also present a second Basenji genome, Wags. The Wags build is of high-quality but not as good as CanFam_Bas, partly because Wags was a male. In total, 64.2% of the missing genes are on the X chromosome, compared to under 16% in the other two individuals. Analysis of long read data mapped onto the CanFam_Bas genome annotation further supports assembly issues as a contributor, with only two autosomal genes lacking any coverage in Wags SMRT data. Equivalently, two other autosomal genes also lacked any coverage in China ONT data, suggesting that even high-quality assemblies can have missing genes in an assembly that may not be biologically lost.

CanFam_Bas is considerably more contiguous and complete than CanFam3.1 (Boxer breed) and slightly more contiguous (Contig N50 37.8 kb vs 20.9 kb) but has slightly lower completeness (BUSCO completeness 92.9% vs 93.0%) than CanFam_GSD [14] (Table 1, Supplementary Table 1, Additional File 2). The ungapped nuclear chromosomes lengths of CanFam_Bas, CanFam3.1, and CanFam_GSD and 2.33 Gb, 2.32 Gb, 2.36 Gb and, respectively. Likely, both assembly quality and breed differences contribute to the 211 missing genes in CanFam3.1. CanFam_GSD had fewer predicted genes missing (95) but a higher percentage (22.0%) also lacking long read coverage. This uncertainty notwithstanding, our analysis highlights the need to consider multiple reference genomes in a pan-genomic approach when a comprehensive analysis is required. These results also highlight the need for considering missing genes carefully on a case-by-case basis as, even with high quality genome assemblies such as those compared in this study, an apparent absence may reflect missing data rather than missing biology.

We noted that the mitochondrial chromosome was missing from the China 1.0 Basenji assembly. An exhaustive search of the genome detected a 33.2 kb region consisting of almost two complete copies of the mitochondrial genome on Chromosome 29 that was not present in the other dog assemblies analysed. This assembly error occurred at a real 4032 bp NUMT. A similar false incorporation of complete mtDNA has been previously reported in the little brown bat [42]. NUMTs are inserted fragments of mtDNA that appear to be present in the nuclear DNA of most, if not all, eukaryotic nuclear genomes [43]. Shotgun genome sequencing data cannot always distinguish NUMTs from mtDNA, which sometimes results in over-stringent removal of NUMTs during genome annotation [43]. Documenting NUMTs in nuclear genome assemblies is important as they have the potential to provide unique insight into population histories and animal well-being [43–45]. Domestic cat nuclear genomes, for example, have 38–76 tandem copies of a 7.9 kb NUMT fragment [46], whilst human NUMTs are polymorphic at 141 loci [43]. NUMT polymorphisms have been used to estimate the age of insertions in human and other primate lineages [44] while five insertions have been implicated in human disease [43, 47]. In total, 212 NUMT blocks (215.1 kb) in 291 fragments (190.5 kb) were detected across the genome, including a previously detected almost full-length (16,577 bp) NUMT on Chromosome 11 [48]. Only one Basenji NUMT was incomplete in the GSD genome. To ease future NUMT comparisons between breeds, we have wrapped up the NUMT discovery and merging methods into an open source tool, NUMTFinder [49].

Sequence analyses estimated six copies of Amy2B in China and five in Wags. Previous studies have shown that Basenjis may have 4–18 copies [6], placing these estimates at the lower end of the range. Short read analysis of 11 other Basenji dogs gave similar copy number estimates of four to seven copies, and 4.38 copies in China. This is consistent with the ddPCR estimate of 4.5 copies, although the uneven coverage of short reads is expected to make this less accurate than the long-read approach. Additional work is needed to establish the source of the differences between ddPCR and read depth estimates. The high variation in Amy2B copy number suggests at least three possible evolutionary histories of the gene in Basenjis. First, Amy2B copies may have differentially accumulated in specific lineages since the divergence from other dog breeds. Second, the ancestral founding population of the modern Basenji may have been polymorphic for Amy2B. Third, it remains possible that the Amy2B region from other breeds have differentially introgressed into the modern breed dog. Basenji-like dogs are depicted in In Egyptian drawings and models dating back to 1550–1900 BC [11]. Possibly, the Basenji is derived from the Abuwtiyuw, which was a lightly built hunting dog with erect ears and a curly tail that died before 2280 BCE [50]. Likely, sequencing the region around Amy2B in Basenjis with a higher copy numbers will aid resolving the alternate hypotheses. This large genetic diversity in Basenji Amy2B copy has important dietary and veterinary implications. Veterinarians frequently recommend a rice-based diet following complex intestinal surgeries and following diarrhoea because it is thought to be bland and easily digestible. However, dogs with low Amy2B copy numbers may not be able to digest rice readily as their serum amylase levels are lower [51].

CanFam_Bas and CanFam_GSD represent two high-quality reference genomes from different breeds. The availability of corresponding Basenji and GSD long read data, provided an excellent opportunity to further investigate the contributions of breed and assembly differences, as real SVs will be represented in the raw read data even if assembly errors have introduced false positives or negatives. Possessing ONT and SMRT data for both GSD and Basenji allows us to overlap the disparate SV call sets to generate a conservative list of SV calls relative to CanFam3.1. This analysis identified over 70,000 SVs in CanFam_Bas relative to CanFam3.1 and over 64,000 SVs in GSD relative to CanFam3.1 (Supplementary Table 7, Additional File 2). There is a high degree of overlap in SVs from GSD and Basenji relative to CanFam3.1 (68% of total basenji SV calls and 77% of GSD SV calls), which highlights potential issues with the current canid reference assembly. Further, each consensus set contains several hundred SVs overlapping annotated exons, highlighting the importance of the selection of appropriate reference genome for analysis of specific genomic regions.

Next, we examined the overlap of the consensus calls for SVs over 100 bp of GSD and Basenji relative to CanFam3.1. We find a high degree of agreement for deletions, with 70.1% of GSD calls and 74.75% of Basenji calls overlapping (Fig. 5b), compared to insertions where 33.3% of basenji calls and 36.4% of GSD calls overlap (Fig. 5c). Basenji and GSD long reads have 25.9 and 19.1% more SVs called when CanFam3.1 is used as the reference, respectively, and the high degree of overlap is consistent with Boxer-specific SVs and/or assembly issues. GSD-specific and Basenji-specific SV calls are also evident.

To get additional insight into breed differences and the influence of reference genome selection, we mapped short read data from 58 breeds onto CanFam_Bas, CanFam3.1, and CanFam_GSD. If assembly differences are dominated by real differences between breeds then we might expect different breeds to map with different efficiencies onto the three genomes. Differences to due quality, on the other hand, should be reflected across all breeds. With some minor exceptions, short read data from the different breeds consistently mapped better onto CanFam_GSD than CanFam_Bas, which was in turn better than CanFam3.1 (Fig. 6a). This is consistent with assembly quality being a dominating factor.

After adjusting for read mapping difference, different reference genomes also produce different SNV and small indel densities for the 58 mapped breeds (Supplementary Fig. 6, Additional File 1). Using a Basenji reference genome consistently identifies more variants than either GSD or Boxer (Fig. 6, Supplementary Fig. 6, Additional File 1). This probably reflects the basal position of Basenji in the breed phylogeny [10]. The basal position of the Basenji makes it useful as a reference for variant analysis as there are clear biases affecting related breeds seen for both the GSD and Boxer reference genomes. On the other hand, CanFam_GSD has slightly higher read mapping levels across breeds, which may provide better total coverage.

Sequencing and assembly efforts are increasingly moving from species reference genomes to breed-specific assemblies, such as those recently published for Great Dane [52], Labrador Retriever [53], and a second GSD [54]. Together, our data suggest that a single high-quality reference should be sufficient for most general analyses, but the generation of breed-specific genomes is likely to be important for canine nutrition and disease studies. The most severe and common ailment in Basenjis was Fanconi Syndrome, in which the renal tubes fail to reabsorb electrolytes and nutrients [55]. Approximately, 10% of Basenjis in North America were affected. In 2011, it was shown that Basenji Fanconi Syndrome is caused by a 370 bp deletion on canine chromosome 3 [56]. To date no other breeds have been recoded with this same deletion, although other mutations can cause the disease. Likely complex diseases in dogs belonging to different clades may have different underlying causes. For example, it remains unclear whether the same suite of mutations causes hip dysplasia in the GSD (New World clade) and the Labrador (Retriever/ Spaniel clade).

Conclusions

Here, we present two high quality de novo Basenji genome assemblies: CanFam_Bas (China, female) and Wags (male). CanFam_Bas offers improved genome contiguity relative to CanFam3.1 and can serve as a representative basal breed in future canid studies. We generate core genomic information for the Basenji that has the potential to inform future studies of population history and aid disease management. We generate high-quality variants (SNVs, small indels, and SVs) relative to CanFam_Bas, CanFam3.1, and CanFam_GSD. We demonstrate the impact that the reference genome makes on both read mapping and variant detection, illustrating the importance of either selecting the appropriate reference genome or employing a pan-genome approach in future canid studies.

Methods

Sequencing and genome assembly of female basenji, China

Sampling

China or Zanzipow Bowies China Girl is an Australian Supreme Champion Kennel Club show dog. Her Australian National Kennel Council registration number is 2100018945. She is bred primarily from Australian lines, with her most recent common ancestor coming from Africa 18 generations previously. She was born on 14 Jan 2016 and is free from all known currently reported diseases.

Sequencing

High molecular weight DNA was extracted from 100 μl of blood using the DNeasy Blood and Tissue kit (Qiagen). For long read (Oxford Nanopore) sequencing, 1 μg of DNA was prepared via the Genomic DNA by Ligation kit (SQK-LSK109) as per the manufacturer’s protocol. Prepared DNA (180 ng) was loaded onto a PromethION (FLO-PRO002) flowcell and sequenced with standard parameters. After 48 h, a nuclease flush was performed, and an additional 115 ng was loaded onto the run. GPU-enabled guppy (v3.0.3) base-calling was performed after sequencing (PromethION high accuracy flip-flop model; config ‘dna_r9.4.1_450bps_hac.cfg’ config).

For short read sequencing whole blood was shipped to BGI, Hong Kong. High molecular weight DNA was extracted, a paired-end library prepared, and the sample run on the BGISEQ-500 sequencing system to generate high quality PE100 data. A total of 767,111,208 clean reads (115.1 Gb) were produced with a lower base call accuracy (Q20) of 95.92%.

Assembly

An overview of the China assembly workflow is given in Supplementary Fig. 1A, Additional File 1. The ONT reads were assembled with the Flye (v2.6-release) assembler [17, 18]. The resulting contigs were polished with ONT reads using four rounds of Racon (v1.4.3) [19] followed by Medaka (v0.10.0) [20] to minimise error propagation. BGI-seq reads were aligned to the polished assembly with BWA-mem (v 0.7.17) [57] and Pilon (v1.23) (diploid) [21] was used for further error correction. A second assembly was performed using Canu assembler (v1.8.0) [58] and error-corrected with two rounds of Arrow polishing [59]. The Flye assembly was considered more contiguous and therefore selected as the core assembly.

Scaffolding

An in situ Hi-C library was prepared [24] from a blood sample from China and sequenced to ~30x coverage (assuming 2.4 Gb genome size). Chromosome-length scaffolding followed the standard DNA zoo methodology (www.dnazoo.org/methods), processing the Hi-C data with Juicer [60] as input for the 3D-DNA pipeline [61]. The resulting candidate scaffolding was manually finished using Juicebox Assembly Tools [22] to produce the final chromosome-length genome assembly. Hi-C matrices are available for browsing at multiple resolutions using Juicebox.js [23] at: https://www.dnazoo.org/post/basenji-the-african-hunting-dog. After scaffolding, all ONT reads were aligned to the assembly with Minimap2 (v2.16) (−ax map-ont) [28] and used by PBJelly (pbsuite v.15.8.24) [25] to fill gaps. BGI reads were re-mapped with BWA-mem (v 0.7.17) and another round of polishing was performed with Pilon (v1.23) (diploid, SNP and indel correction) [21] .

Final clean-up

The Pilon-polished genome underwent a final scaffold clean-up to generate a high-quality core assembly following Field and colleagues’ processing of Canfam_GSD [14]. Scaffolds were reoriented and assigned to CanFam3.1 chromosomes [15] using PAFScaff (v0.3.0) [14] (Minimap2 v2.16 mapping) [28]. Diploidocus (v0.9.0) [27] was used to screen contamination, remove low-coverage artefacts and haplotig sequences, and annotate remaining scaffolds with potential issues as described in [14]. ONT reads were mapped onto the assembly using Minimap2 (v2.17) (−ax map-ont --secondary = no) [28] and read depth summaries calculated with BBMap (v38.51) pileup.sh [62]. Any scaffolds with median coverage less than three (e.g., less than 50% of the scaffold covered by at least three reads) were filtered out as low-coverage scaffolds. Single-copy read depth was estimated using the modal read depth of 35X across the 55,578 single copy complete genes identified by BUSCO (v3.0.2b) [33]. This was used to set low-, mid- and high-depth thresholds at 8x, 25x and 68x for PurgeHaplotigs v20190612 [63] (implementing Perl v5.28.0, BEDTools v2.27.1 [64, 65], R v3.5.3, and SAMTools v1.9 [66]). PurgeHaplotig coverage was adjusted to exclude gap regions and scaffolds filtered as in [14] for haplotigs and assembly artefacts (scaffolds with 80% + bases in the low/haploid coverage bins and 95% + of their length mapped onto another scaffold by PurgeHaplotigs) or low coverage artefacts (remaining scaffolds with 80% + low coverage bases). Remaining scaffolds were further classified based on read depth profiles: scaffolds with < 20% diploid coverage and 50% + high coverage were marked as probable collapsed repeats; scaffolds with dominant diploid coverage and > 50% match to another scaffold were marked as a possible repeat sequences [14].