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

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. 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. 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.


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][4][5][6][7][8][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 highresolution 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.

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).

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)  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.

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

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 Boxerwe 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 Can-Fam_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 over- 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.

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). Largescale 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).

Discussion
In this manuscript we present a reference-quality assembly of a female Basenji (China), which we designate Can-Fam_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, Can-Fam3.1, 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][44][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, NUMT-Finder [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 highquality 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 Can-Fam3.1 (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 Can-Fam3.1. 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, Can-Fam3.1, 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 highquality 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, Can-Fam3.1, 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.

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. 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%.

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 Mini-map2 (v2.16) (−ax map-ont) [28] and used by PBJelly (pbsuite v. 15.8.24) [25] to fill gaps. BGI reads were remapped with BWA-mem (v 0.7.17) and another round of polishing was performed with Pilon (v1.23) (diploid, SNP and indel correction) [21] .

Correction of mitochondrial insertion into chromosome 29
NUMT analysis identified a 33.2 kb region consisting of almost two complete copies of the mitochondrial genome, not present in other dog genome assemblies. GABLAM (v2.30.5) [29] was used to confirm that this region was also absent from the Canu assembly of China v1.1. ONT reads that mapped onto both flanking regions of the 33.2 kb putative NUMT were extracted and reassembled with Flye (v2.7.1) [17,18]. The new NUMT was approx. 2.8 kb long. To avoid repeated problems with mitochondrial reads mis-polishing the sequence, a subset of 4.82 M ONT reads (72.7 Gb,~30X) was extracted and mapped onto the assembled region with Minimap (v2.17) [28]. Reads mapping to at least 5 kb of the assembled region including some immediate flanking sequence were extracted (66 reads, 1.50 Mb) and polished with one round of Racon (v1.4.5) [19] (−m 8 -x − 6 -g − 8 -w 500) and Medaka (v0.7.1) [20] (model r941_prom_high). The polished NUMT region was mapped on to the Chromosome 29 scaffold with GABLAM (v2.30.5) [29] (blast+ v2.9.0 [67] megablast) and stretches of 100% sequence identity identified each side of the NUMT. The mtDNA sequence between these regions of identity was replaced with the re-assembled NUMT sequence.

DNA methylation calling
China's blood DNA methylome libraries were sequenced on the Illumina HiSeq X platform (150 bp, PE), generating 336 million read pairs and yielding 14x sequencing coverage. Sequenced reads were trimmed using Trimmomatic [73] and mapped to the China v1.0 genome reference using WALT [74] with the following parameters: -m 10 -t 24 -N 10000000 -L 2000. The mappability of the MethylC-seq library was 86%. Duplicate reads were removed using Picard Tools (v2.3.0). Genotype and methylation bias correction were performed using MethylDackel with additional parameters: minOpposite-Depth 5 --maxVariantFrac 0.5 --OT 10,140,10,140 --OB 10,140,10,140. The numbers of methylated and unmethylated calls at each CpG site were determined using M e t h y l D a c k e l ( h t t p s : / / g i t h u b . c o m / d p r y a n 7 9 / MethylDackel). Bisulphite conversion efficiency was 99.71%, estimated using unmethylated lambda phage spike-in control. Nanopore reads were aligned to the generated reference genome using Minimap2 v2.17 [28], and CpG methylation sites were called with f5c v0.6 [75], which is an accelerated version of nanopolish [76]. Methylation frequency was then collated for each CpG site.

Comparison of DNA methylation calling by PromethION and MethylC-seq
Average MethylC-seq and PromethION DNA methylation for 1 kb genomic bins, UMRs and LMRs were calculated using the overlapRatios R function. Scatterplots were generated using comparisonplot R function.
Sequencing and genome assembly of male basenji, wags  (Fig. 1b) were generated on the Pacific Biosciences Sequel instrument (V2 chemistry) to approximately 45x genome coverage based on a genome size estimate of 2.5 Gb. An overview of the China assembly workflow is given in Supplementary Fig. 1B, Additional File 1. All SMRT sequences were assembled with the HGAP4 algorithm, a Falcon based assembly pipeline available through the SMRT Link interphase (SMRT Link v5.0.1.9585) [77]. The assembly was then error corrected with the original SMRT sequences using the Arrow error-correction module [77]. Additional polishing of the assembly for residual indels was done by aligning 32x coverage of Illumina data and the Pilon algorithm [21]. Chromosomal level scaffolds were generated with the same DNA source using the Proximo™ Hi-C genome scaffolding software (Phase Genomics Inc) and finalized by alignment to the CanFam3.1 reference.

Locus copy number estimation
Copy numbers for specific assembly regions were calculated using Diploidocus (v0.10.0) (runmode = regcnv) [27]. For each animal, long reads were mapped onto the assembly with Minimap2 (v2.17) (no secondary mapping) [28] and the modal read depth across single-copy Complete genes identified by BUSCO v3 [33] (laura-siatherian_db) calculated using SAMTools (v1.9) [66] mpileup. This set the expected single-copy read depth, X SC . Copy number for a region, N reg was estimated by dividing the mean read depth across that region, X reg , by X SC . The variance and standard deviation of the estimate was calculated using X reg for all single copy BUSCO genes. For Wags (a male), genes on the X chromosome were excluded from this analysis.

Amylase copy number
The copy number of the beta amylase gene Amy2B was calculated using Diploidocus (v0.10.0) (runmode = regcnv) [27] using a modification of the single locus copy number estimation (above) to account for multiple copies of the gene in the assembly. First, the AMY2B protein sequence from CanFam3.1 (UniprotKB: J9PAL7) was used as a query and searched against the genome with Exonerate (v2.2.0) [78] to identify assembled copies of the Amy2B gene. A full-length (14.8 kb) Amy2B gene repeat was also extracted from the CanFam_GSD assembly and mapped onto the assembly with Minimap2 (v2.17) (−x asm5) [28]. Estimated N reg values were converted into a number of copies by multiplying by the proportion of the query found covered by that region. The total genome Amy2B copy number was then calculated as the summed copy number over each hit. To further investigate the robustness of the method and improve the Amy2B copy number estimate in CanFam_Bas, analysis was repeated with ONT reads at least 5 kb in length and at least 10 kb in length. These reads should be less susceptible to poor mapping at repeat sequences, but at a cost of reduced coverage. A simpler approach was also applied to short read data for China and eleven Basenji dogs with data available on SRA (Supplementary Table 4, Additional File 2). Reads were mapped onto the China genome with BWA mem [57] and sequencing depth calculated across the annotated AMY2B genes from GeMoMa (Chromosome 6: 46,943,564-46,950, 679 and 46,928,705-46,935,822) using samtools v1.11 [66]. The mean sequencing depth per base per AMY2B copy was then compared to the mean sequencing depth across the genome, calculated with BBMap (v38.51) pileup.sh [62]. AMY2B copy number was estimated using the AMY2B:genome read depth ratio.

Genome annotation
Each genome was annotated using GeMoMa [38] (v1.6.2beta, mmseqs2 [80] v5877873cbcd50a6d954607fc2df1210f8c2c3a4b) homology-based gene prediction and nine reference organisms as in Field et al. [14]. To make a fair comparison of the influence of genome quality and completeness on annotation, CanFam3.1 was annotated with the same pipeline. Annotation of CanFam_GSD using the same pipeline was obtained from Field et al. [14].

Annotation summary and quality assessment
Annotation summary statistics and the longest protein isoform per gene were generated with SAAGA (v0.4.0) [81]. Annotation completeness was estimated using BUSCO v3 [33] (laurasiatherian, n = 6253, proteins mode), run on a reduced annotation consisting of the longest protein per gene. To check for truncated or fragmented protein predictions, predicted proteins were mapped onto the Quest For Orthologues reference dog proteome [39] with mmseqs2 v [80]. The best protein hit for each gene was used to calculate a protein length ratio (length of predicted protein / length of reference protein). Percentage coverage of query and hit proteins was calculated with mmseqs2 v [80]. A reciprocal search identified the closest predicted protein for each reference protein. Any reciprocal best hits were marked as predicted orthologues.

Annotation copy number and coverage analysis
Predicted copy number was calculated for every proteincoding gene in CanFaBas using the method described above. In addition, assembly coverage for each CanFam_ Bas gene was calculated for Wags, CanFam3.1 and CanFam_GSD. Each genome was split into 1 Mb tiled fragments and mapped onto CanFam_Bas with Mini-map2 (v2.17) (−ax asm5 -L) [28]. Each BAM file was used for Diploidocus (v0.10.0) regcnv analysis with a single-copy read depth of 1x. In addition, SMRT reads from Wags and ONT reads from the GSD were mapped onto CanFam_Bas with Minimap2 (v2.17) (−-secondary = no -L -ax map-pb or -ax map-ont) [28] and the standard predicted copy number calculation applied. Genes with zero coverage were marked 0n. Other genes were binned according to coverage: for mapped assemblies, coverage was rounded to the nearest integer; for long read mapping, coverage was rounded to the nearest 0.5n. Genes with greater than zero but less than 50% coverage were assigned to 0.5n. Any genes exceeding a rounded coverage of 2n were grouped as "3+".

Whole genome assembly comparisons
Whole genome synteny analysis were performed for the main chromosome scaffolds using the D-GENIES [30] web portal.

Short read mapping, SNV / small indel detection
Representative Illumina data was identified from [85]. All Sequence Read Archive (SRA) IDs associated with the DBVDC bioproject on NCBI (SRP144493) were downloaded along with their metadata and reduced to 126 samples representing the biggest sequencing run (no. bases) per annotated breed. These 126 samples were used for initial read mapping and variant calling (Supplementary Table 9, Additional File 2). Following removal of unknown/mixed/village dog samples, canids other than domestic dogs, and duplicates the remaining breeds were mapped onto those used by Parker et al. [10]. In total, 58 breeds in the significantly monophyletic clades (greater than 70% bootstrap) designated by Parker et al. [10] were considered for analyses (Supplementary Table 8, Additional File 2).
SNVs and small indels were called from the Illumina reads of the 58 representative breeds against three reference genomes (Basenji China v1.0, CanFam3.1, and Can-Fam_GSD). All Illumina reads were downloaded from the short read archive using the SRA toolkit (v2.10.4) [86]. All samples were analysed using a modified version of an existing variant detection pipeline [87]. Briefly, the pipeline employs BWA (v0.7.17-r1188) for read alignment [88], SAMTools (v1.9) [66] and picard (v2.4.1) (http://broadinstitute.github.io/picard) for binary alignment map (BAM) file preprocessing, and Genome Analysis ToolKit (v3.6) (GATK) for calling SNVs and small indels [89]. The workflow follows GATK bestpractices using default parameters throughout. Additionally, SRA reads from each reference genome were aligned using BWA and SAMTools and a consensus variant list generated using an approach described previously [90]. The consensus variant lists for each reference genome were utilized in GATK's BaseRecalibrator step as the '-knownSites' argument to serve as the required variant truthset. Variant alignment statistics were generated using SAMTools flagstat. Joint variant calls were generated for the larger dataset of 126 samples relative to the three reference genomes and the total number of SNVs and small indels for individual samples tabulated. Variants for the Basenji and CanFam_GSD were annotated with gene model predictions generated using GeMoMa (see above) while CanFam variants were annotated with ENSEMBL gene annotations v100 for CanFam3.1. Read mapping statistics for each sample were calculated using SAMTools to remove secondary mapping and then summarized with BBTools (v38.51) pileup.sh [62]. Numbers were then converted into relative values for each reference by averaging the score for each breed over the three reference genomes and then calculating the difference from the mean. Breeds were considered individually, but mean values for each clade were also calculated. Three clades are of particular importance as they are closely related, or include, the three reference genome assemblies. The Asian Spitz clade is considered closely related to the Basenji. This clade contained the Alaskan Malamute, Shar-Pei, Shiba Inu and Tibetan Mastiff. The European Mastiff Clade contained the Boxer, Bull Terrier, Cane Corseo, Great Dane, Mastiff and Rhodesian Ridgeback. The New World clade contained the Berger Picard, Chinook and German Shepherd. One-way ANOVA's were employed to detect significant differences between groups. As the same short read samples were examined relative to the three reference genomes statistical significance was set to be P < 0.01.
Additional file 1: Supplementary Fig. 1. The workflow used to construct the de novo genomes. Supplementary Fig. 2. KAT kmer analysis of China assembly. Supplementary Fig. 3. Identification of putative regulatory elements in basenji's genome using whole-genome bisulphite sequencing. Supplementary Fig. 4. Predicted copy number of BUSCO Complete genes for three dogs based on long-read read depth. Supplementary Fig. 5. CanFam_Bas nuclear mitochondrial DNA (NUMT) coverage. Supplementary Fig. 6. Comparative short read mapping and single nucleotide variant calling for 58 dog breeds versus three reference genomes: CanFam_Bas, CanFam_GSD (GSD) and CanFam3.1 (BOX).