Genome sequences of two diploid wild relatives of cultivated sweetpotato reveal targets for genetic improvement

Sweetpotato [Ipomoea batatas (L.) Lam.] is a globally important staple food crop, especially for sub-Saharan Africa. Agronomic improvement of sweetpotato has lagged behind other major food crops due to a lack of genomic and genetic resources and inherent challenges in breeding a heterozygous, clonally propagated polyploid. Here, we report the genome sequences of its two diploid relatives, I. trifida and I. triloba, and show that these high-quality genome assemblies are robust references for hexaploid sweetpotato. Comparative and phylogenetic analyses reveal insights into the ancient whole-genome triplication history of Ipomoea and evolutionary relationships within the Batatas complex. Using resequencing data from 16 genotypes widely used in African breeding programs, genes and alleles associated with carotenoid biosynthesis in storage roots are identified, which may enable efficient breeding of varieties with high provitamin A content. These resources will facilitate genome-enabled breeding in this important food security crop.

S weetpotato, Ipomoea batatas (L.) Lam. (2n = 6x = 90), is an important food crop ranking seventh globally with 106.6 million tons produced in 2014 (FAOSTAT 2014). Sweetpotato provides a rich source of carbohydrates, dietary fiber, vitamins, and micronutrients, is low in fat and cholesterol 1 , and due to its resilience and adaptability, it serves an important role in food security for subsistence farmers in sub-Saharan Africa (SSA). Indeed, efforts by the International Potato Center (CIP) to replace the dominant β-carotene lacking white-fleshed varieties in SSA with provitamin A-rich orange-fleshed sweetpotato (OFSP) 2 were considered a breakthrough achievement as vitamin A deficiency affects more than 40% of children under 5 years old in SSA and is a leading cause of blindness and premature death 3 . Awarding of the World Food Prize in 2016 to four scientists who pioneered biofortification with OFSP highlights the significance of these efforts and the importance of OFSP in shifting human health outcomes.
Sweetpotato was domesticated in tropical America at least 5000 years ago in a region between the Yucatan Peninsula of Mexico and the Orinoco River in Venezuela 4 , and introduced to Europe and Africa from South America by Europeans in the early 16th century and later to the rest of the world 5 . Whereas sweetpotato has long been thought to have been introduced to Oceania by Polynesians in pre-Columbian times 5 , this hypothesis has recently been challenged 6 . Sweetpotato is one of only two food crops (the other being water spinach, I. aquatic Forssk.) among 1600 species in the morning glory family (Convolvulaceae). The wild diploid species I. trifida (Kunth.) G. Don (2n = 2x = 30) has been widely reported as the closest relative of cultivated sweetpotato 6,7 , but there are two hypotheses regarding the origin of the cultivated hexaploid sweetpotato: an autopolyploidization within an I. trifida progenitor population 6,7 or an alloautopolyploidization involving both I. trifida and another diploid wild relative, possibly Ipomoea triloba L. (2n = 2x = 30) 4 . Some accessions of I. trifida are sexually compatible with I. batatas, though this is rare and, unlike I. triloba, commonly produces unreduced pollen associated with polyploidization 8,9 . Observations of a predominance of polysomic inheritance have been interpreted as support for the autopolyploidization hypothesis 10 . However, frequent tetrasomic recombination has been observed in allopolyploids, such as cultivated peanut (Arachis hypogaea) 11 . Some molecular data implicate I. trifida as a close, but not direct ancestor, and provide no evidence for I. triloba as contributing to the hexaploid genome of sweetpotato 7,12 . Roullier et al. 7 speculate that sweetpotato could share a common ancestor with I. trifida and/or tetraploid I. batatas yet have multiple origins with at least two autopolyploidization events from a single progenitor species. Munoz-Rodrıguez et al. 6 recently used targeted gene capture to analyze hexaploid I. batatas and wild relatives in an attempt to resolve the origin of the hexaploid sweetpotato, but their findings have yet to be tested using whole-genome sequence data.
The improvement of sweetpotato varieties across the world faces major constraints due to the lack of knowledge of the genetic, molecular, and physiological basis of key agronomic traits of this critical food crop. The sweetpotato genome is large (1.6 Gb) 13 and complex due to polyploidy and a high degree of heterozygosity, challenging efforts to generate a high-quality de novo genome assembly. While an assembly of I. trifida was released in 2015, it was highly fragmented 14 , and thus its use as a reference for genome-enabled breeding efforts in sweetpotato is limited. Likewise, a recently reported haplotype-resolved genome assembly of I. batatas 15 is incomplete and contains a considerable amount of redundancy and misassemblies, including erroneous assembly of haplotypes. To provide high-quality reference genome resources and further investigate the relationship between hexaploid sweetpotato and its diploid relatives, we generated high quality, chromosome-scale genome assemblies of I. trifida and I. triloba, resequenced 16 cultivars and landraces widely used in African breeding programs, and performed a suite of comparative genomic analyses. Our findings illustrate the utility of highquality diploid I. trifida and I. triloba genomes for genomeenabled breeding and research in this important food security crop.

Results
Genome assemblies of I. trifida and I. triloba. We sequenced the genomes of I. trifida NCNSP0306 and I. triloba NCNSP0323, two diploid species closely related to the hexaploid sweetpotato ( Fig. 1 and Supplementary Method 1). A total of 104.8 and 144.1 Gb high-quality cleaned Illumina paired-end and mate-pair reads were generated for I. trifida and I. triloba, respectively, representing 199× and 291× coverage of their genomes (Supplementary Table 1). Based on the 17-mer depth distribution of the Illumina reads ( Supplementary Fig. 1 Table 2).
To construct pseudomolecules, a high-density genetic map for I. trifida was generated using an F 1 population derived from a cross between two heterozygous I. trifida lines, M9 and M19 (Supplementary Method 3 and Supplementary Fig. 2). In all, 461 I. trifida scaffolds with a cumulative size of 373.4 Mb (80.8% of the assembly) were anchored to 15 linkage groups (LGs), of which 196 (66.3%) were oriented (Supplementary Table 3). The I. trifida genetic map was also used to anchor 216 I. triloba scaffolds (443.3 Mb; 96.8% of the assembly) to the 15 LGs (Supplementary  Table 3).
To evaluate the quality of the assemblies, RNA-Seq reads from a set of developmental tissues were aligned to their cognate assemblies; 85.3% and 91.4% of the reads aligned to the I. trifida and I. triloba genomes, respectively (Supplementary Data 1). The completeness of the assemblies in terms of gene content was assessed with BUSCO 16 . Of the core conserved plant genes, 93.9% and 94.6% were complete in the I. trifida and the I. triloba genome assemblies, respectively, while another 1.9% and 1.6%, respectively, were fragmented. Collectively, these results demonstrate that our genome assemblies are of high quality. Comparative genomics and lineage-specific family expansion. We performed comparative analyses using 43,296 protein sequences derived from a transcriptome assembly of I. batatas "Beauregard" (Supplementary Method 4) and the predicted proteomes from I. trifida, I. triloba, and seven other representative plant species, including I. nil (Japanese morning glory), Solanum lycopersicum (tomato), Solanum tuberosum (potato), Vitis vinifera (grapevine), Arabidopsis thaliana (model eudicot), Oryza sativa (rice), and Amborella trichopoda (basal angiosperm). A total of 274,278 proteins from the 10 species (79.3% of the input sequences) clustered into 19,901 orthologous groups, of which, 3269 were present only in the four Ipomoea species, and 1680 were specific to the Batatas complex (I. batatas, I. trifida, and I. triloba) and absent in I. nil (Fig. 2a).
From orthologous gene clustering with other angiosperm proteomes, 73 gene families (orthologous groups) were significantly expanded in I. trifida and/or I. triloba (Supplementary Data 3). This lineage-specific expansion gene set was enriched with biological processes including "response to stress", "cell death", "response to external stimulus", and "response to biotic stimulus" (P < 0.0001, Fisher's exact test; Supplementary Data 4). It is worth noting that two orthologous groups containing genes encoding sporamin proteins were expanded in both I. trifida and I. triloba. Sporamin, a Kunitz-type trypsin inhibitor (KTI), is a major storage protein in sweetpotato storage roots, and is analogous to patatin in potato which plays an important role in storage, defense, and development 17,18 . Based on the presence of the Kunitz soybean trypsin inhibitor domain (PF00197), we identified 94 and 83 putative KTIs in I. trifida and I. triloba, respectively, whereas 41, 36, and 16 were present in I. nil, potato, and tomato, respectively, and only seven, five, one, and two were found in Arabidopsis, grape, rice, and Amborella, respectively. Phylogenetic relationships of KTIs in these nine species revealed a subgroup of KTIs specific to I. trifida and I. triloba, as well as groups of expanded KTIs in potato ( Supplementary Fig. 3); suggestive of lineage-specific adaptation events in these root and tuber crops. In particular, two I. batatas sporamin genes, orthologous to the I. trifida genes itf10g10920 and itf01g01870 that are within the Ipomoea-lineage-specific expansion, were upregulated during storage root development in "Beauregard", with expression levels 13-74-fold higher in storage roots compared to fibrous roots (Supplementary Data 5).
Ipomoea whole-genome triplication. We identified syntenic blocks within the I. trifida and I. triloba genomes and inferred whole-genome multiplication events based on the distributions of pairwise synonymous substitution rates (Ks) of paralogous genes. The Ks distributions in both I. trifida and I. triloba showed a peak at 0.65 (Fig. 2b). A substantial proportion of the extant Ipomoea genomes occurred in triplicated blocks (23.5%, 34.8%, and 28.0% in I. trifida, I. triloba, and I. nil, respectively), and these proportions are higher than that observed in the tomato genome (6.0%) (Fig. 2c), which bears a signature of whole-genome triplication (WGT) 19 . We further compared the genomes of I. trifida, I. triloba, and I. nil, respectively, with the grape genome. A total of 72.9% of the I. trifida gene models are in syntenic blocks corresponding to one grape region. These grape regions collectively cover 79.9% of the grape gene space, among which 36.1% have three orthologous regions in I. trifida, 27.1% have two, and 16.8% have one. Similar results were obtained for I. triloba and I. nil (Supplementary Table 4). Based on the syntenic blocks involving single grape genome segments, each of the I. trifida, I. triloba, and I. nil genomes can be partitioned into three nonoverlapping "subgenomes" (Supplementary Fig. 4). Our data indicate that WGT occurred in an ancient ancestor of the Ipomoea lineage. While both the tomato and Ipomoea WGT events were indicated by peaks of Ks near 0.65, the Ks distribution for tomato, and I. trifida orthologous gene pairs is shifted to the right of the paralog-pair distributions for either species (Fig. 2b), indicating that independent WGTs occurred in the ancestry of Ipomoea and Solanaceae after their divergence. Based on the evolutionary rate of 7.0 × 10 −9 substitutions/site/year reported in both Solanaceae 20 Fig. 5), suggesting limited large-scale interchromosomal rearrangements over the last 3.6 million years. A robust gene expression atlas of developmental and stress/ hormone treatments (Supplementary Method 4) was used to investigate the fate of gene copies following the ancient WGT event. Genes retained in triplicate or duplicate in the I. trifida and I. triloba genomes were significantly enriched (P < 0.0001, Fisher's exact test) for transcripts that exhibited up-regulation in response to abiotic and biotic stressors, suggesting WGT may play a role in adaptive stress responses. WGT genes were also enriched with gene ontologies suggestive of regulatory functions including "DNA binding", "transcription factor activity", "protein binding", "kinase activity", and "translation regulator activity" (P < 0.0001, Fisher's exact test; Supplementary Data 4). For example, two sweetpotato knotted-like homeobox genes, IBKN2 (orthologous to itf01g32840) and IBKN3 (itf01g32840) known to be associated with storage root development 22 , were duplicated in the ancestral Ipomoea WGT event. In 'Beauregard', these two genes were expressed at higher levels in storage roots as compared with fibrous roots and were upregulated during the development of storage roots (Supplementary Fig. 6), suggesting that the WGT contributed to additional gene copies that function in storage root development.
Utility of diploid genomes as references for sweetpotato. Yang et al. 15 recently presented an 836-Mb haplotype-resolved genome assembly of I. batatas "Taizhong6". We assessed the quality of the "Taizhong6" assembly using multiple approaches and found the assembly was incomplete and contained numerous misassemblies, limiting its use as a reference for hexaploid sweetpotato (Supplementary Method 5 and Supplementary Figs. 7-12). To evaluate the utility of our two Ipomoea diploid genomes as references for hexaploid sweetpotato, we generated~60× wholegenome sequence data for the hexaploid African landrace "Tanzania" using the 10× Genomics Chromium system (Supplementary Method 6) and aligned the linked reads to the I. trifida and I. triloba genome assemblies separately;~83.5% of the reads could be aligned to both assemblies while~7.7% of the reads did not align (Fig. 3a). Approximately, 5.4% and 3.4% of the reads aligned solely to the I. trifida and I. triloba genome assemblies, respectively (Fig. 3a). Based on the alignment scores,~57.7% of the reads aligned better to the I. trifida genome than the I. triloba genome, while~31.9% of the reads aligned better in the reciprocal direction (Fig. 3b).     Molecules were reconstructed from the linked reads of "Tanzania" using their alignments to the two diploid genome assemblies. Approximately, 4.3 and 4.8 million "Tanzania" molecules of at least 10 kb were identified from alignments to the I. trifida and I. triloba genome assemblies, respectively, covering~409 Mb (83%) and~396 Mb (86%) of the I. trifida and I. triloba genomes. We identified 390,303 regions in "Tanzania" that were homologous to both diploid genomes, covering~215 Mb (44%) and~231 Mb (50%) of the I. trifida and I. triloba genome assemblies, respectively (Fig. 3c). We further identified regions that were specific to I. trifida (~7.0 Mb) or I. triloba (~9.9 Mb) genomes, suggesting that the hexaploid sweetpotato genome contains sequences that are uniquely shared with either of these two diploid species (Fig. 3c). Collectively, our results suggest that the two diploid genome assemblies can serve as robust reference sequences for hexaploid sweetpotato as suggested by the high percentages (>90%) of hexaploid reads aligned to either I. trifida or I. triloba. The I. trifida genome is a slightly better reference, while the I. triloba genome shares unique features with the hexaploid genome and can serve as a complement to the I. trifida genome.
We further analyzed the frequency of the I. trifida-type and I. triloba-type alleles in "Tanzania". A total of 7.6 million SNP and indel variants were identified between the I. trifida and the I. triloba genome assemblies, of which, approximately 5.6 million variant sites had at least 10× depth of coverage from the 10× Genomics data of "Tanzania". The majority of these sites (~73.9%) were nonpolymorphic in the "Tanzania" genome comprising both I. trifida-specific (~46.2%) and I. triloba-specific (~27.7%) homozygous sites ( Supplementary Fig. 13a). This observation indicated that both I. trifida-like and I. triloba-like progenitors had contributed to the "Tanzania" genome, and the I. trifida-like progenitor had contributed approximately twice as much as the I. triloba-like progenitor. The similar allele depth distribution for the I. trifida-specific and I. triloba-specific homozygous sites ( Supplementary Fig. 13b) in the hexaploid "Tanzania" genome suggested that it was genetic drift that mainly caused the fixation of one allele in all six copies, rather than absence of one genotype in one of the subgenomes leading to four copies of the I. trifida-like allele and two copies of the I. trilobalike allele. Interestingly, we seldom found "Tanzania" molecules specific to one progenitor type (<1%; i.e., those with all reads supporting only the I. trifida-specific or I. triloba-specific alleles). Instead, the reads supporting the alleles from different homeologous subgenomes were interwoven within 10× molecules, suggesting recombination between the subgenomes. The proportion of I. trifida specific alleles within molecules peaked around 0.6 ( Supplementary Fig. 13c), suggesting the approximate amount of contribution of the I. trifida-like progenitor genome to the hexaploid sweetpotato genome.
Facilitating genome-enabled breeding for subsistence farmers. The Mwanga diversity panel (MDP) contains 16 sweetpotato cultivars, breeding lines and landraces selected from germplasm sourced from different areas across Uganda and a few selected introductions. The MDP accessions have been extensively used as parents in African sweetpotato breeding programs 23 . Previous work using 31 simple sequence repeat (SSR) markers separated the 16 accessions into two genetic groups, population Uganda A and population Uganda B 23 . To revisit their phylogeny and genetic diversity, and to reveal genes and alleles associated with agronomic traits, we resequenced these 16 accessions and aligned the reads to the I. trifida and I. triloba reference assemblies (Supplementary Data 6). Due to the highly heterozygous nature of the subgenomes within hexaploid sweetpotato, especially within intergenic regions, accurate variant calling was restricted to genic regions, and a total of 6,138,575 and 6,722,808 highconfidence variant loci were identified in the I. trifida and I. triloba reference assemblies, respectively, from these 16 accessions (Supplementary Table 5). For both reference genomes, approximately 95% of these loci were bi-allelic. Genotypes of the MDP accessions were consistent with a highly heterozygous genome in hexaploid I. batatas, with heterozygous calls at >50% of SNP loci in each accession. The proportions of allele dosage were roughly equivalent among the MDP accessions using either reference genome, at roughly a 3:2:1 ratio for simplex to duplex to triplex dosage classes (Supplementary Fig. 14a, b).
SNP-based genetic clustering of the MDP accessions was similar, but inconsistent with their population membership determined previously using SSR markers 23 (Fig. 4a and Supplementary Fig. 14c), and our results, derived using over one million bi-allelic SNPs with no missing calls for all MDP accessions, represent an improved delineation of their genetic relationships due to the higher resolution available with this highdensity, robust polymorphic marker set. Accessions from population A formed two separate and highly supported clusters, whereas population B accessions did not cluster well, with "Magabali", "NASPOT 11", and "New Kawogo" forming a wellsupported clade. Principal component analysis (PCA) of the 16 MDP accessions further supported their relationships (Supplementary Fig. 15a), and population structure analysis revealed admixture in most of the 16 accessions ( Supplementary Fig. 15b), which was not surprising as several of the lines used in the MDP crossing block can trace their parentage to landraces obtained from farmers in Uganda.
Median read depth was relatively stable from chromosome to chromosome for most accessions, but increased read depth on chromosome 1 initially suggested 6× +1 aneuploidy in "Huarmeyano" and "Mugande", while decreased read depth on chromosomes 8 and 11 suggested 6× −2 aneuploidy in "NASPOT 5" (Fig. 4b and Supplementary Fig. 16). There was no substantial decrease or increase in SNP density for the chromosomes with putative 6× +1 or 6× −2 changes, respectively, suggesting that these chromosomal aberrations in read depth are not likely due to differences in similarity to the reference genomes. Thus, we implemented a statistical test to detect aneuploidy based on read depth (Supplementary Method 7). The test indicated that "NASPOT 5" had a~0.9 loss of chromosome 8 (p < 1.5e-36 for ≥5.13 copies) as well as a~0.6 loss of chromosome 11 (p < 4.7e-78 for ≥5.39 copies), and "Huarmeyano" had a~0.3 gain on chromosome 1 (p < 1e-20 for ≤6.3 copies) and "Mugande" a~0.5 gain on chromosome 1 (p < 3e-97 for ≤6.52 copies) (Supplementary Fig. 17). Consistent with our read mapping results, counts of chromosomes in root tip cells confirmed that "NASPOT 5" is a double monosomic line with 88 chromosomes, whereas "Mugande" contains 90 chromosomes (as do "Beauregard" and "Tanzania"), and thus, is not an aneuploid ( Fig. 4c and Supplementary Fig. 18). Structural variation in the form of copy number and presence/absence variation have been reported previously in a wide range of crop species and have been associated with phenotypic variation [24][25][26] . Analyses in potato, another vegetatively propagated polyploid crop, revealed extensive structural variation [27][28][29][30] , including presence/absence variation of sequences up to 575 kb in length that impacts transcript dosage. The discovery of aneuploidy in cultivated sweetpotato may present an extreme form of structural variation that clearly would affect transcript dosage and consequently, phenotypic variation.
Orange flesh in sweetpotato is correlated with high β-carotene content, and cultivation and consumption of OFSP has been promoted in SSA to help alleviate vitamin A deficiency 31 .
However, other traits such as yield, dry matter content, disease and insect resistance, and stress tolerance are important in development of consumer-accepted varieties. The MDP is composed of orange-and white-fleshed accessions (Supplementary Data 6), and to enable stacking high β-carotene with other agronomic traits, SNPs associated with flesh color in the MDP were identified. The identified SNPs occurred in homologs of genes encoding key carotenoid biosynthetic enzymes, including phytoene synthase (PSY), phytoene desaturase (PDS), ζ-carotene isomerase (Z-ISO), and lycopene β-cyclase (LCYB) (Fig. 5a, b, Supplementary Data 7 and Supplementary Method 8). The PSY (itf03g05110), PDS (itf11g08190) and Z-ISO (itf04g12320) genes were of particular interest as they were upregulated in orangefleshed 'Beauregard' storage roots as compared to other root types (Fig. 5c). In contrast, the LCYB (itf04g32080) gene had reduced expression in storage roots as compared to fibrous roots, suggesting it may not be involved in conferring orange flesh in "Beauregard" storage roots. Furthermore, three SNPs in PSY that were significantly associated with orange flesh in the MDP showed biased expression of the "orange" alleles in "Beauregard" storage roots at 40 and/or 50 days after transplanting (DAT) (Supplementary Fig. 19), suggesting that the basis of increased carotenoid accumulation in "Beauregard" is expression-based. One SNP associated with orange flesh in PDS was weakly significant for biased expression of the "orange" allele at 20 DAT yet not significant in 50 DAT storage roots, suggesting that expression of specific PDS alleles prior to storage root initiation might be important in the development of orange flesh. Together, our analysis identified logical control points for the accumulation of β-carotene in OFSP, and these loci provide targets for markerassisted selection of orange-fleshed varieties.
Origins and evolution of the Batatas complex. Phylogenetic inferences attempting to describe the evolutionary history of the Batatas complex have been highly incongruent 4,6,32,33 , yet no study to date has explicitly tested for hybridization or reticulate evolution within the complex. Rather than a strictly bifurcating tree, the true evolutionary relationships among species or populations within the Batatas complex may be best understood as a network with reticulations 34,35 . To test the hypothesis of recent or ancient hybridization in the Batatas complex, we performed targeted sequence capture on 490 genes that are single copy in the I. trifida and I. triloba genomes as well as genome assemblies of other core eudicots as described in Amborella Genome Project 36 14 ), and twelve polyploid I. batatas accessions. Haplotype diversity was assessed for each gene within each of the I. batatas assemblies and only genes with a single haplotype (i.e., nulliplex) or two haplotypes that were sister in the gene trees were included in phylogenetic analyses. The coalescence-based species-tree estimation did not account for reticulations in the phylogeny, but nonetheless recovered six distinct lineages within the Batatas complex ( Fig. 6a and Supplementary Fig. 20). Ipomoea triloba accessions formed a clade, whereas relationships could not be resolved within a clade including all I. trifida and cultivated sweetpotato (I. batatas) accessions, and an unnamed diploid accession. Another unnamed diploid accession was recovered as sister to all other members of the Batatas complex aside from I. splendor-sylvae (Fig. 6a). While there was no overlap in accessions used in this study and a recently published analysis of Batatas complex species relationships based on 307 exon-capture assemblies and the whole plastid genome 6 , both analyses infer early divergence of ancestral lineages for I. splendor-sylvae and the remainder of the Batatas complex, followed by divergence of lineages leading the last common ancestors of accessions in the I. batatas + I. trifida clade and the I. triloba + I. lacunosa + I. cordatotriloba + I. tenuissima clade.
Reticulations in the phylogeny were tested using pseudolikelihood method for inferring phylogenetic networks 37 . The PhyloNet 37 analysis was run using a single exemplar for polyploid I. batatas together with all of the diploid Batatas complex accessions used in the species-tree analysis. The likelihood of a network with two reticulations (Fig. 6b) was significantly better than a tree with the same samples estimated without reticulations (p < 0.001, likelihood ratio test; Supplementary Fig. 20d), but otherwise the relationships among cultivated sweetpotato (using 'Tinian' as an exemplar), I. trifida and I. triloba remained unchanged. The reticulation giving rise to cultivated sweetpotato exhibited evidence of a wide ancestral cross. However, since the taxonomy of accession PI 552786 collected from Bolivia is uncertain, the results do not clearly favor conspecific parents (autopolyploidy) or interspecific hybridization (allopolyploidy) within the I. trifida + I. batatas clade as giving rise to polyploid I. batatas. The deeper reticulation in the PhyloNet network suggests that extensive gene flow between divergent ancestral populations gave rise to the clade including I. triloba, I. lacunosa, I. cordatotriloba and I. tenuissima. With respect to I. batatas "Tinian", the inference that parental lineages were genetically divergent (Fig. 6b) was consistent with results of mapping the I. batatas "Tanzania" genomic reads to the I. trifida and I. triloba reference genomes and allele frequency analysis ( Fig. 3 and Supplementary Fig. 20). This finding underscores the continuum between autopolyploidy and allopolyploidy in outcrossing species.

Discussion
Sweetpotato is an important food security crop with a highly recognized potential to alleviate hunger, vitamin A deficiency, and poverty in SSA. We sequenced the genomes of two diploid relatives of sweetpotato, I. trifida and I. triloba, and demonstrated that these high-quality assemblies can be used as robust references to facilitate sweetpotato breeding. While I. trifida is the more closely related diploid to hexaploid sweetpotato, and is a better reference for sweetpotato, the I. triloba genome can be a complementary reference sequence.
Many species in the Batatas complex are able to hybridize with one another, and hybridization seems to be important in the evolution of Batatas complex populations [38][39][40] . Some diploid species within the complex have been hypothesized as hybrid species, and sweetpotato is thought to have arisen through either allopolyploidization 4 , autopolyploidization 6 , or both 10 . Furthermore, microsatellite data provide evidence supporting on-going gene flow between sympatric populations of I. cordatotriloba and I. lacunosa 41 . The observation that many species in the Batatas complex appear morphologically similar and show evidence of hybridization may be the result of ongoing or ancient hybridization. Results from read mapping and phylogenetic analyses suggest that cultivated sweetpotato is derived from a cross between divergent parents but their taxonomy is equivocal.
Comparative analyses using the diploid genome sequences revealed an ancient WGT event specific to the Ipomoea genus. We identified gene families specifically expanded in I. trifida and/ or I. triloba associated with adaptation and stress resistance in roots including sporamin genes, which could be a prerequisite for storage root development in the hexaploid sweetpotato. Genome resequencing of 16 sweetpotato cultivars and landraces used in African breeding programs allowed refinement of the genetic relationships among these key accessions, an understanding of dosage distribution within cultivated sweetpotato, and identification of genes and alleles associated with high β-carotene content, highlighting how genomic tools can enable more efficient improvement of sweetpotato. We anticipate that these resources will become increasingly valuable for the global sweetpotato breeding and genomic community as we seek to further improve this critical food crop using genome-enabled marker-assisted breeding methods.

Methods
De novo genome assembly. For I. triloba NCNSP0323, high-quality cleaned Illumina paired-end and mate-pair reads (Supplementary Method 1) were assembled into scaffolds with SOAPdenovo2 (ref. 42 ) and gaps in the resulting scaffolds were filled with GapCloser in the SOAPdenovo2 package. For I. trifida NCNSP0306, cleaned paired-end and mate-pair reads were assembled into scaffolds using Platanus 43 (v1.2.1). The error-corrected PacBio long reads were used to further fill gaps in the scaffolds and to connect scaffolds using PBJelly 44 (v15.8.24). Pilon 45 (v1.13) was used to improve the assembly by correcting base errors, fixing misassemblies, and further filling gaps. Potential contamination from microorganisms was detected by aligning the assemblies to the NCBI nonredundant nucleotide (nt) database using BLASTN with an e value cutoff of 1e−5. Scaffolds with more than 90% of their length similar to only bacterial or virus sequences were considered contaminants and removed. Finally, scaffolds contained within other scaffolds with sequence identity >99% and coverage >99% were removed.
Genome annotation. Custom repeat libraries were individually created for I. trifida and I. triloba by combining the putative repeat libraries predicted from MITE-Hunter 46 (v2011) and RepeatModeler (http://www.repeatmasker.org/; v1.0.8). Protein-coding genes were removed from the repeat library using ProtExcluder 47 . The Repbase (v20150807) repeats for green plants (Viridiplantae) were then added to each library to create a final custom repeat library for each species. The pseudomolecules for I. trifida and I. triloba were repeat masked with the respective repeat library using RepeatMasker (http://www.repeatmasker.org/; v4.0.6).
For gene prediction, AUGUSTUS 48 (v3.1) was trained on the soft-masked assemblies using the leaf RNA-Seq alignments. Gene models were predicted with AUGUSTUS using the hard-masked assemblies and refined with PASA2 (v2.0.2) (ref. 49 ) using the genome-guided transcript assemblies from each tissue as transcript evidence (Supplementary Method 1). Two rounds of gene prediction comparison were performed and gene models PASA identified as merged, but unable to split, were manually inspected and split as necessary. A third round of gene prediction comparison was performed to refine the structure of the manually curated models. The final output from PASA2 is the working set of gene models. Expression abundances for each gene model were determined based on the RNA-Seq read alignments using Cufflinks2 (ref. 50 ) (v.2.2.1). A high-confidence gene model set was constructed from the working gene model set by removing partial gene models and gene models with an internal stop codon, a hit to a transposable element, or an FPKM of 0 across the RNA-Seq libraries used for the annotation.
Functional annotations for the high-confidence gene models were assigned by comparing their protein sequences against the Arabidopsis proteome (TAIR 10), Pfam (v29), and the Swiss-Prot databases. Proteins that only matched a hypothetical Arabidopsis gene model and had no matches in the other databases were annotated as conserved hypothetical, while proteins with no matches in any of the databases were annotated as hypothetical. Lineage-specific gene family expansion was inferred from orthologous groups using the software CAFE 52 (v3.0). I. batatas was excluded from the analysis because its transcriptome assembly did not represent the complete gene content. The random gene birth and death rates were estimated using a maximum likelihood method 52 across the species tree composed of the nine species. Orthologous groups with accelerated rate of expansion in I. trifida and/or I. triloba were determined with a branch-specific Viterbi p value cutoff of 0.01. Given that CAFE may fail to provide accurate lambda estimates for gene families with large differences in size among species, we also considered gene families expanded if they contained more than ten genes in either I. trifida or I. triloba and at least 50% more genes in these two species than any of the other investigated species. Fisher's exact test followed by a Bonferroni multiple test correction was used to determine gene enrichment within 96 gene ontology terms as well as for tissue specific, biotic and abiotic stress conditions.
Synteny and whole-genome duplication analysis. To identify syntenic regions within and between genomes, protein sequences of I. trifida, I. triloba, I. nil, tomato, and potato were self-aligned and aligned with each other using BLASTP. High-confidence collinear blocks were determined using MCScanX 53 with an e value cutoff of 1e−10. Ks values of homologous pairs were calculated using the Yang-Nielsen algorithm implemented in the PAML package 54 . Divergence time (T) was calculated as T = Ks/2r, where r is the evolutionary rate (7.0 × 10 −9 substitutions/site/year), according to those reported in the Solanaceae 20 (6.96 × 10 −9 ) and Arabidopsis 21 (7.0 × 10 −9 ).
10× Genomics data analysis. We generated 10× Genomics sequences for hexaploid sweetpotato cultivar "Tanzania" (Supplementary Method 6). To reconstruct "Tanzania" molecules sequenced by the 10× platform, we grouped the reads by barcodes. For each group, we then identified clusters of linked reads, which were mapped to the same region in the reference genome. Two consecutively mapped reads were considered linked (i.e., sequenced from the same "Tanzania" molecule) if their barcodes were identical and the alignment distance between them was ≤10 kb. Each linked read cluster was regarded as a span of a "Tanzania" molecule. Only clusters spanning at least 10 kb in the reference genome were retained. It should be noted that, given the aforementioned method for the molecule reconstruction, each "Tanzania" molecule represented a homologous region between the "Tanzania" genome and the diploid genome that was used as the reference. If a "Tanzania" molecule is homologous to both diploid genomes, i.e., the two linked read clusters for a "Tanzania" molecule obtained from using the two diploid reference genomes are identical, then we considered it to be a three-way homology meaning the homologous region was shared between the "Tanzania" genome and the two diploid genomes.
Variant detection and phylogeny of hexaploid sweetpotatoes. Whole genome sequence reads from the 16 MDP accessions (Supplementary Method 7) were cleaned using Cutadapt 55 (v1.11) with a base quality threshold of 20 and a minimum length of 200 bases after trimming. Reads were aligned to the I. trifida and I. triloba genome assemblies using BWA-MEM 56 (v0.7.12) in paired-end mode with default parameters. Picard tools (v2.7.1; http://broadinstitute.github.io/picard) were used to remove duplicate read pairs and alignments with a MAPQ score less than 30. Alignments were refined using the GATK IndelRealigner 57 (v3.6.0). Alignments to the same reference assembly from all individuals were analyzed together for variant and genotype calling at regions overlapping genes only. FreeBayes 58 (v1.1.0) was used with an assumed ploidy of 6, "haplotype-length" of 0, minimum base quality of 20, and a minimum of 2 reads and 10% of the reads in at least one single individual supporting an alternate allele. Variant records were filtered using the "vcffilter" function in vcflib (v1.0.0; https://github.com/vcflib/vcflib) with a minimum variant quality score of 40, "SAF" > 1, "SAR" > 1, "RPR" > 1, and "RPL" > 1. Individual genotypes were filtered using a minimum depth of 18, maximum depths based on average depth across genic regions plus 2 standard deviations for each sample, and a minimum of 5% of reads supporting the least frequent allele in a b  Fig. 6 Phylogenetic relationships among members of the Batatas complex. a Phylogenetic relationships among Batatas complex accessions inferred using the species-tree method ASTRAL-II. Bootstrap support values are shown next to the branches. Relationships among I. trifida and I. batatas accessions were not resolved. b Phylogenetic network of the Batatas complex including pseudolikelihood estimates for hybridization events in the ancestor of red, green and yellow clades as well as in the ancestry of cultivated sweetpotato. Blue lines indicate the two-reticulation events inferred in PhyloNet. Below the phylogenetic network are −log-likelihood scores for the tree generated in PhyloNet under the zero-reticulation model (Supplementary Fig. 20d) and the network estimated under the two-reticulation model. Color bars at the right of each tree denote the six inferred major lineages a genotype call. Complex variants were decomposed to SNPs and indels using the "vcfallelicprimitives" function in vcflib and variant records were removed if all the genotypes containing the alternate allele failed the above quality filters. All analyses were performed using these filtered variant and genotype calls, with additional processing as described.
Phylogenetic analysis was performed using bi-allelic SNPs with no missing data for any accessions, with Rphylip 59 (v0.1-23). Nei's genetic distances were calculated using 'gendist' and a neighbor-joining tree was constructed using 'neighbor'. Bootstrap analysis was done using 1000 replications created with 'boot. phylo' from the R package, ape 60 (v4.1). PCA analysis was performed using PLINK 61 (v1.9) and population structure analysis using STRUCTURE 62 (v2.3.4). The most likely number of clusters (K) was estimated by calculating delta K 63 . STRUCTURE analyses were run 20 times for each K value ranging from 1 to 10, using 8,000 randomly selected SNPs with an admixture model. The subgroup memberships of accessions were determined based on 10,000 iterations.
Phylogenomic analysis of species in the Batatas complex. We performed targeted sequence capture on 490 single-copy genes for fifteen accessions in the Batatas complex and computationally extracted target exon sequences from available genome assemblies for I. trifida (NCNSP0306, ITR and ITRk 14 ) and I. triloba (NCNSP0323), and from shotgun reads from 18 I. batatas accessions (Taizhong6 15 , Tanzania, and 16 MDP accessions) (Supplementary Method 9). Multiple sequence alignment for each single-copy gene family was performed using PRANK 64 , and alignments were filtered using Gblocks 65 to remove poorly aligned regions. Gene trees were estimated separately for each gene alignment in RAxML with 100 bootstraps 66 (v8.0.0). For each gene, samples exhibiting more than a single-hapotype or an accession-specific clade in the gene tree were removed (Supplementary Method 9). After filtering, 11 hexaploid I. batatas accessions had fewer than 30 retained single-hapotype genes and were removed from further analysis. A majority of the remaining 12 I. batatas accessions were retained in 85 gene trees. These single-copy/single-haploid gene trees were included in ASTRAL-II 67 (v4. 10.12) and SVDQuartets 68 estimates of relationships among the Batatascomplex species and accessions while accounting for incomplete lineage sorting.
PhyloNet 37 was used to infer a phylogenetic network while accounting for incomplete lineage sorting under zero, one, two, and three reticulation scenarios using the InferNetwork_MPL option and gene tree from RAxML as the input. Five independent analyses of each reticulation scenario were carried out to best traverse the complex parameter space. The phylogenetic network with the best likelihood score, which was significantly different from the likelihood of a zero-reticulation network (i.e., the ASTRAL-II and SVDQuartets trees) was chosen as the most likely network given the gene trees. PhyloNet is computationally intensive, so only the 'Tinian' accession was included as an exemplar for I. batatas.

Data availability
The genome assemblies and raw genome reads of I. trifida and I. triloba have been deposited into GenBank BioProject under accessions PRJNA428214 and PRJNA428241, respectively. Raw 10x Genomics, genome resequencing and RNA-Seq reads are available in the National Center for Biotechnology Information Sequence Read Archive under accessions SRP161954, SRP162006, SRP132113, SRP132112, SRP162112, SRP162110, and SRP162021. The pseudomolecules, genome annotation, expression abundances, BLAST server, and Jbrowse instance are available in the Sweetpotato Genomics Resource (http://sweetpotato.plantbiology.msu.edu). In addition, the genome assemblies, annotated genes, expression abundances, and variants of the MDP are available via the Dryad Digital Repository (https://doi.org/10.5061/dryad.b9m61cg). Results of phylogenetic analysis and the associated multiple sequence alignments were available in the Github repository (https://github.com/laeserman/Wu-et-al.