Main

Barley remains dated to the dawn of agriculture have been found at several archaeological sites1,2. In addition to indications that barley was an important food crop, recent excavations have fuelled speculation that beverages from fermented grains may have motivated early Neolithic hunter–gatherers to erect some of humankind’s oldest monuments3,4. Moreover, brewing beer may also have played a role in the eastward spread of the crop after its initial domestication in the Fertile Crescent5,6.

Since 2012, both genetic research and crop improvement in barley have benefited from a partly ordered draft sequence assembly7. This community resource has underpinned gene isolation8,9 and population genomic studies10. However, these and other efforts have also revealed limitations of the current draft assembly. The limitations are often direct consequences of two characteristic genomic features: the extreme abundance of repetitive elements, and the severely reduced frequency of meiotic recombination in pericentromeric regions11. These factors have limited the contiguity of whole-genome assemblies to kilobase-sized sequences originating from low-copy regions of the genome. Thus, a detailed investigation of the composition of the repetitive fraction of the genome—including expanded gene families—and of the distribution of targets of selection and crop improvement in (genetically defined) pericentromeric regions has been beyond reach.

Here we present a map-based reference sequence of the barley genome including the first comprehensively ordered assembly of the pericentromeric regions of a Triticeae genome. The resource highlights a conspicuous distinction between distal and proximal regions of chromosomes that is reflected by the intranuclear chromatin organization. Moreover, chromosomal compartments are differentiated by an exponential gradient of gene density and recombination rate, striking contrasts in the distribution of retrotransposon families, and distinct patterns of genetic diversity.

A chromosome-scale assembly of the barley genome

We adopted a hierarchical approach to generate a high-quality reference genome sequence of the barley cultivar Morex, a US spring six-row malting barley. First, a total of 87,075 bacterial artificial chromosomes (BACs) were sequenced, mainly using Illumina paired-end and mate-pair technology and assembled individually from 4.5 terabases of raw sequence data12,13,14 (Supplementary Note 1). In a second step, overlaps between adjacent clones15 were detected and validated by physical map information16, a genetic linkage17 and a highly contiguous optical map18 to construct super-scaffolds composed of merged assemblies of individual BACs (Table 1 and Extended Data Table 1). This increased the contiguity as measured by the N50 value (the scaffold size above which 50% of the total length of the sequence was included in the assembly) from 79 kb to 1.9 Mb. Scaffolds were assigned to chromosomes using a population sequencing (POPSEQ) genetic map17. Finally, we used three-dimensional proximity information obtained by chromosome conformation capture sequencing19,20,21 (Hi-C) to order and orient BAC-based super-scaffolds (Supplementary Note 2 and ref. 22). The final chromosome-scale assembly of the barley genome consists of 6,347 ordered super-scaffolds composed of merged assemblies of individual BACs, representing 4.79 Gb (~ 95%) of the genomic sequence content, of which 4.54 Gb have been assigned to precise chromosomal location in the Hi-C map (Table 1). Mapping of transcriptome data and reference protein sequences from other plant species to the assembly identified 83,105 putative gene loci including protein-coding genes, non-coding RNAs, pseudogenes and transcribed transposons (Fig. 1, Extended Data Fig. 1, Extended Data Table 2 and Supplementary Note 3). These loci were filtered further and divided into 39,734 high-confidence genes (with four different sub-categories) and 41,949 low-confidence genes on the basis of sequence homology to related species (Methods and Supplementary Note 3.4). Moreover, we predicted 19,908 long non-coding RNAs (Supplementary Note 3.7) and 792 microRNA precursor loci (Supplementary Note 3.8). The high co-linearity between the Hi-C-based pseudomolecules and linkage and cytogenetic maps22 as well as the conserved order of syntenic genes in pericentromeric regions compared with model grass Brachypodium distachyon (Extended Data Fig. 2a) corroborated the quality of the assembly. Extrapolating from a set of conserved eukaryotic core genes23, we estimate that the predicted gene models represent 98% of the cultivar Morex barley gene complement (Extended Data Fig. 2b).

Table 1 Assembly and annotation statistics
Figure 1: Characteristics of genomic compartments in barley chromosomes.
figure 1

a, The distribution of genomic features in 4 Mb windows is plotted along chromosome 1H. Analogous panels for the other chromosomes are found in Extended Data Fig. 5a. The left column in the legend refers to the background shading in the top panel; the right column indicates the colour code for lines in both panels. CDS, predicted coding sequences; cM, centimorgans. b, Enrichment of Gene Ontology (GO) terms in genomic compartments. Coloured rectangles indicate enrichment factors ranging from −2 (dark blue) to 2 (dark red). Numbers inside the rectangles indicate −log10-transformed P values.

PowerPoint slide

Organization of chromatin

Barley has served as a model for traditional cytogenetics11; but relating chromosomal features to unique sequences has been challenging, requiring the cloning of repeat-free probes24. The reference sequence allowed us to employ the Hi-C data to interrogate the three-dimensional organization of chromatin in the nucleus. As in other eukaryotes20,25,26, the spatial proximity of genomic loci as measured by Hi-C link frequency is highly dependent on their distance in the linear genome (Fig. 2a). However, we observed an elevated link frequency at distances above 200 Mb and a pronounced anti-diagonal pattern in the intrachromosomal Hi-C contact matrices (Fig. 2b and Extended Data Fig. 3a), indicating an increased adjacency of regions on different chromosome arms. We interpret this pattern as reflective of the so-called Rabl configuration27 of interphase nuclei, where individual chromosomes fold back to juxtapose the long and short arms, with centromeres and telomeres of all chromosomes clustering at opposite poles of the nucleus (Fig. 2c and Supplementary Fig. 2.2). Fluorescence in situ hybridization (Fig. 2d) supported this hypothesis. Principal component analysis of the intrachromosomal proximity matrix showed that the first three principal components cumulatively explained ~70% of the variation and differentiated (1) distal from proximal regions, (2) interstitial from both distal and proximal regions and (3) the long arms from the short arms (Fig. 2f and Extended Data Fig. 4a). A linear model taking into account the genomic distance between two loci, as well as their relative distance from the centromere, accounted for 79% of the variation (Extended Data Fig. 4b) in the intrachromosomal proximity matrix at 1 Mb resolution.

Figure 2: Chromosome conformation capture analysis.
figure 2

a, Distance-dependent decay of contact probability. b, Intrachromosomal contact matrix. The intensity of pixels represents the normalized count of Hi-C links between 1 Mb windows on chromosome 1H on a logarithmic scale. c, Schematic model of the Rabl configuration of interphase chromosomes. Centromeres and telomeres are presented by red and green circles, respectively. d, Leaf interphase nucleus of barley. Chromatin was stained blue with 4′,6-diamidino-2-phenylindole (DAPI). Fluorescence in situ hybridization was performed with probes specific for centromeres (red) and telomeres (green). Scale bar, 5 μm. e, Interchromosomal contact matrix. The intensity of pixels represents the normalized count of Hi-C links between 1 Mb windows on chromosomes 1H (x axis) and 2H (y axis) on a logarithmic scale. A principal component analysis of the normalized contact matrix at 1 Mb resolution of chromosome 1H was conducted. f, The first and second eigenvectors are plotted against each other. Each point represents a 1 Mb window. Closer proximity to the centromere is indicated by a darker colour. Windows from the short and long arms are coloured blue and red, respectively.

PowerPoint slide

Contacts between loci on different chromosomes followed a similar pattern (Fig. 2e and Extended Data Fig. 3b): a prominent cross pattern supporting a juxtaposition of long and short arms. In contrast to intrachromosomal matrices, contact probabilities between loci on, for instance, the short arm of one chromosome are equal for loci on both the short and the long arm on another chromosome having the same relative distance to the centromeres: that is, facing each other in the interphase nucleus. We also observed a higher contact frequency between telomere-near regions, as has been observed in Arabidopsis25.

To test whether pairs of homologous chromosomes are positioned closer to each other than to non-homologues, we performed diploid Hi-C28 on leaf tissue from F1 hybrids between the cultivars Morex and Barke, and assigned the resultant Hi-C links to the haplotypes of both inbred parents by mapping reads to a diploid reference. We did not observe any preferential interaction between homologues. Rather, contacts between the maternal and paternal copies of the same chromosome occurred as frequently as between non-homologues (Extended Data Fig. 4c).

We conclude that the frequency with which loci juxtapose in three-dimensional space is predominantly determined by their position in the linear genome. This is in sharp contrast to the organization of chromatin in human nuclei where two compartments corresponding to open and closed chromatin domains are evident at megabase resolution20, but is consistent with cytogenetic mapping of histone marks associated with heterochromatin in large, repeat-rich genomes29.

The genomic context of repetitive elements

Large plant genomes consist mainly of highly similar copies of repetitive elements such as long terminal repeat (LTR) retrotransposons and DNA transposons30,31. Our hierarchical sequencing strategy reduced the algorithmic complexity of assembling a highly repetitive genome from short reads. Instead of resolving complex repeat structures on the whole-genome level, we reconstructed the sequences of 100–150 kb BACs. This allowed us to disentangle nearly identical copies of highly abundant repetitive elements, as evidenced by the good representation of both mathematically defined repeats and retrotransposon families (Extended Data Fig. 2c, d). Homology-guided repeat annotation with a Triticeae-specific repeat library32 identified 3.7 Gb (80.8%) of the assembled sequence as derived from transposable elements (Table 1, Fig. 1a and Extended Data Table 3), most of which were present as truncated and degenerated copies, with only 10% of mobile elements intact and potentially active.

Median 20-mer frequencies were used to partition the seven barley chromosomes into three zones (Fig. 1 and Extended Data Fig. 5a), reminiscent of the three compartments of wheat chromosome 3B33. The distal zone 1 was characterized by an enrichment of low-copy regions, a high gene content and frequent meiotic recombination. Zone 2, occupying the interstitial regions of chromosomes, had the highest 20-mer frequencies and intermediate gene density. Surprisingly, the abundance of repetitive 20-mers decreased in the proximal zone 3, where older mobile elements with diverged, and thus unique, sequences predominated (Fig. 1). The three zones also differed in the composition of the gene space (Extended Data Table 2b and Supplementary Note 3). For example, genes involved in defence response and reproductive processes were preferentially found in distal regions, while proximal regions contained more genes related to housekeeping processes, such as photosynthesis and respiration, compared with other parts of the genome (Fig. 1b).

Transposable element groups exhibited pronounced variation in their insertion site preferences (Fig. 3a and Extended Data Fig. 5b). On a global scale, most miniature inverted-repeat transposable elements and long interspersed elements were found in gene-rich distal regions, as has been reported in other grass species34,35. By contrast, zone 3 was populated by Gypsy retrotransposons, while Copia elements favoured zones 1 and 2. These differences in the relative abundance of retrotransposon families were reflected by distinct distributions of functional domains. For example, sequences encoding the chromodomain (PF00385) are concentrated in the vicinity of the centromere and may be involved in the target specificity through incorporation in the integrase of Gypsy elements36 (Fig. 3a and Extended Data Fig. 5b).

Figure 3: The genomic context of repetitive elements.
figure 3

a, Abundance of key genomic features, different transposon superfamilies and common Pfam domains across chromosome 1H. Analogous panels for the other chromosomes are found in Extended Data Fig. 5b. The colour scale of the heatmaps ranges from blue (0) to yellow (maximum across all chromosomes per track). Minimum and maximum values are indicated to the right of each track. MITEs, miniature inverted-repeat transposable elements; LINEs, long interspersed elements; fl, full-length; PR, protease; CH, chromodomain; RT, reverse transcriptase; NBS, NB-ARC; Pkin, protein kinase. b, Transposable elements up- and downstream of genes. Coding sequences of high-confidence genes were used as anchor points. Transposable element composition was determined 10 kb up- and downstream of each gene. The x axis indicates the position relative to the gene, while the y axis indicates how many genes had a transposable element of the respective superfamily at the respective position in their upstream/downstream region.

PowerPoint slide

At a local scale, different types of elements also occupy different niches in the proximity of genes (Fig. 3b). Mariner transposons preferably reside within 1 kb up- or downstream of the coding regions of genes, while Harbinger and long interspersed elements are found further away. The observed distribution of different types of transposable elements around genes may reflect selective pressures, allowing only the smallest elements, namely Mariners, to be tolerated closest to genes. Intriguingly, Helitrons as well as elements of the Harbinger superfamily have a clear preference for promoter regions, while long interspersed elements have a preference for downstream regions (Fig. 3b). At greater distances from genes, large elements such as LTR retrotransposons and CACTA elements dominate.

Expansion of gene families

The barley reference sequence enabled us to disentangle complex gene duplications that may shed light on gene family expansion specific to barley or the Triticeae. A total of 29,944 genes belonged to families with multiple members (Fig. 4a and Supplementary Note 4.1). Gene families expanded in barley were tested for over-representation of Gene Ontology37 terms compared with sorghum, rice, Brachypodium and Arabidopsis. Among the most significant results were terms related to defence response and disease resistance (NBS-LRR and thionin genes), as well as thioredoxin genes (Supplementary Note 4.1).

Figure 4: Expansion of agronomically important gene families.
figure 4

a, OrthoMCL clustering of the barley high-confidence gene complement with B. distachyon, rice, sorghum and Arabidopsis thaliana genes. Numbers in the sections of the Venn diagram correspond to numbers of clusters (gene groups). The first number below the species name denotes the total number of proteins that were included into the OrthoMCL analysis for each species. The second number indicates the number of genes in clusters for a species. b, Phylogenetic tree of 68 full-length α-amylase protein sequences derived from amy genes identified in the genomes of barley, hexaploid wheat, B. distachyon, rice, sorghum and maize. Each wheat subgenome was considered separately to facilitate the comparison of gene copy numbers and duplication events across species. Note that for the amy4 subfamily, two to three genes per genome were identified in all genomes. These genes are located on distinct chromosomes and hence most probably did not originate from tandem gene duplications. While most species further contain only a single amy3 gene copy per genome, moderate copy number extension was observed in sorghum and rice where a potential tandem gene duplication resulted in two amy3 gene copies. Three genes of the amy2 subfamily were identified on chromosome 7H in barley and on chromosomes 7A, 7B, 7D in wheat. No similar copy number extension was observed in B. distachyon, Sorghum bicolor or Oryza sativa. In maize, two amy2 genes were identified. The amy1 subfamily shows the highest level of copy number extension. Tandem duplications are present in sorghum and rice. Two to three full-length genes were identified per genome in hexaploid wheat on group 6 chromosomes and five full-length amy1 genes on chromosome 6H and unanchored scaffolds in barley. Notably four of these barley genes share 99.8–100% sequence identity on protein and nucleotide level, indicating very recent duplication events. T. aestivum, Triticum aestivum; Z. mays, Zea mays. c, Expression of the SWEET11 gene subfamily in the developing barley grains. Left, expression profiles of SWEET11a and SWEET11b as determined by quantitative real-time PCR (qPCR) on total RNA isolated from micro-dissected developing grains. Right, localization of SWEET11a and SWEET11b expression in cross-section of immature seeds by RNA in situ hybridization. Hybridizations with sense probes are shown as negative controls in Extended Data Fig. 7a. Scale bars, 100 μm.

PowerPoint slide

In the following, we focused on a detailed analysis of gene families having particular importance for malting quality. Germinating barley grains possess high diastatic power: that is, the combined ability of a complex of enzymes to mobilize fermentable sugars from starch. Key diastatic enzymes include α-amylases. The genome of barley cultivar Morex contains 12 α-amylase (amy) family sequences (Supplementary Note 4.2 and Extended Data Table 4a), which can be classified into four subfamilies38. Gene duplication events have occurred in the subfamilies amy1 and amy2 (Fig. 4b), located on chromosomes 6H and 7H, respectively. The existence of these duplications had been speculated earlier, but could not be analysed further because of high sequence similarity between the copies. The reference assembly contained five full-length amy1 subfamily genes, four of which, here designated as amy1_1a–d, shared >99.8% identity at the nucleotide level including introns. Locus-specific PCR confirmed earlier suggestions39,40 of multiple, highly similar amy1_1 genes (Extended Data Fig. 6 and Supplementary Note 4.2). Given the relevance of α-amylase activity to the brewing process, the high variability of the amy1_1 multiple gene locus (Extended Data Fig. 6) observed in landraces and elite lines, including modern malting cultivars, is remarkable.

The accumulation of fermentable carbohydrates in the grain depends on the transfer of sugars from maternal tissue into the developing seeds. In contrast to the two routes of nutrient transfer in rice seeds—the nucellar projection and nucellar epidermis—delivery of assimilates into barley grains occurs predominantly via the nucellar projection41 and requires active transporters. The family of SUGARS WILL EVENTUALLY BE EXPORTED TRANSPORTER (SWEET) transmembrane proteins mediating sugar efflux42 consists of 23 members in barley (Extended Data Table 4b and Supplementary Note 4.3). There is a small extension of the sugar-transporting SWEET11, SWEET13, SWEET14 and SWEET15 subfamilies, with two or more genes for each subgroup compared with only a single orthologue in rice and Arabidopsis (Extended Data Table 4b). Duplication of SWEET11 was most likely followed by neofunctionalization as evidenced by divergent expression patterns. Both SWEET11a and SWEET11b were highly expressed in maternal seed tissue, but differed in the distribution of expression domains (Fig. 4c and Extended Data Fig. 7). Genes encoding a family of vacuolar processing enzymes, which are essential for programmed cell death in maternal tissue43 and starch accumulation in the grain (Supplementary Note 4.3 and V.R., unpublished observations) showed a similar expansion in barley (Extended Data Table 4c), pointing to the central role of the nucellar projection for grain filling in the Triticeae.

These examples of genes involved in sugar transport and metabolism illustrate that the high-quality reference genome sequence can serve as a springboard for the in-depth analysis of the evolutionary history of gene duplications, their relation to morphological and physiological innovations, and their impact on crop performance.

Molecular diversity and haplotype analysis

To explore how the new barley genome assembly could be exploited for genetics and breeding, we generated exome sequence data from 96 European elite barley lines, half with a spring growth habit, half with a winter one (Supplementary Table 5.1). We investigated the extent and partitioning of molecular variation within and between these groups using 71,285 single-nucleotide polymorphisms (SNPs). Plotting diversity values in 100 SNP windows both in linear order (Fig. 5a) and according to physical distance (Fig. 5b) revealed marked contrasts in the levels and distribution of diversity both within and between gene pools. In spring types, extensive regions on chromosomes 1H, 2H and 7H were virtually devoid of diversity, as was a large region on 5H in the winter gene pool. For these chromosomes, this results in a single gene-pool-specific haplotype across the extensive pericentromeric regions. Chromosomes 3H, 4H and 6H maintain higher diversity across these regions owing to the presence of multiple similarly extensive haplotypes. This is even more evident when diversity is plotted on a physical scale (Fig. 5b). We presume that the lack of observed variation in elite germplasm is a signature of intense selection during breeding for different end-use sectors (principally malting versus feed barley), and the virtual absence of allelic re-assortment during meiosis owing to restricted recombination in the pericentromeric regions.

Figure 5: Distribution of genetic diversity across the barley genome.
figure 5

Ninety-six elite barley cultivars, including 48 from the winter gene pool (blue line) and 48 from the spring gene pool (red line), were used. Diversity (unbiased heterozygosity, y axis) is plotted as the rolling average of 100 adjacent SNPs along each chromosome. For improved visualization, all chromosomes have been normalized to a standard length. a, Patterns of diversity on chromosomes 1H–7H (top to bottom). The distance between each SNP has been normalized (that is, this does not show genetic distance). The number of SNPs included on each chromosome is given at the bottom right of each plot. b, The same diversity values normalized according to physical distance. Extensive peri-centromeric regions of very low diversity in the spring gene pool are highlighted in green and low diversity in the winter gene pool in purple. Regions with similar levels of diversity in both gene pools are highlighted in orange. Coloured dots show the position of eight loci previously identified as being differentiated between the winter and spring gene pools.

PowerPoint slide

Crosses between spring and winter barleys are rarely performed as they are considered to disrupt the gene-pool-specific gene complexes required for general performance (such as phenological adaptations) and end-use quality. Contrasting local patterns of diversity outside the pericentromeric regions therefore also most likely reflect the outcome of selection within alternative gene pools. We explored this further by comparing diversity in eight characterized genes whose variant alleles are important for conditioning barley’s seasonal growth habit (Supplementary Note 5). Of the eight genes, HvCEN is uniquely ‘locked’ in the pericentromeric region of chromosome 2H where alternative alleles at a single SNP confer both differences in days-to-heading44 and strong latitudinal differentiation10. The extensive pericentromeric haplotype in spring barleys (Fig. 5) may stem from selection for this single HvCEN SNP. While strong selection for other favourable alleles locked in the same region in spring barley cannot be ruled out, the virtual absence of recombination severely restricts exploitation of diversity across the entire region. Despite our focus here on life-history traits, strong selection for other traits mapping to pericentromeric regions45,46, including good malting quality in the spring gene pool on chromosomes 1H and 7H, would probably also reduce diversity in these regions. Interestingly, we are unaware of any phenotypic trait in the winter gene pool that would result in strong selection for a single pericentromeric haplotype on chromosome 5H.

We next explored patterns of linkage disequilibrium across the entire genome. As expected for two highly inbred and elite crop gene pools, we observed extensive linkage disequilibrium on all chromosomes in both spring and winter barleys (Extended Data Fig. 8). The number of discrete haplotype blocks in this germplasm set varied from 86 to 161 per chromosome (Extended Data Fig. 8). Surprisingly, the two-row spring gene pool, generally considered to be narrowest owing to intense selection for malting quality, exhibited a greater number of haplotype blocks than the winter lines for most chromosomes.

Discussion

To assemble a highly contiguous reference genome sequence for barley, we combined hierarchical shotgun sequencing, a strategy previously used for assembling large and complex plant genomes33,47, with novel technologies such as optical mapping18 and chromosome-scale scaffolding with Hi-C21. The latter technology was key to resolving the linear order of sequence scaffolds in pericentromeric regions. We anticipate the adoption of Hi-C-based genome mapping in other Triticeae species, such as bread and durum wheat and their wild relatives. Now that the quality of whole-genome shotgun assemblies is on a par with map-based assemblies48,49, we believe that the barley genome project will be one of the last such efforts to follow the laborious BAC-by-BAC approach.

The barley reference genome sequence constitutes an important community resource for cereal genetics and genomics. It will facilitate positional cloning, provide a better contextualization of population genomic datasets and enable comparative genomic analysis with other Triticeae in non-recombining regions that have been inaccessible to analysis of gene collinearity until now. The exciting methodological advances in sequence assembly and genome mapping have enabled even large and repeat-rich genomes to be unlocked48,50 and hold the promise of constructing reference-quality genome sequences, not only for a single cultivar, but also for representatives of major germplasm groups.

Methods

No statistical methods were used to predetermine sample size. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment.

Sequencing and assembly of individual BAC clones

Barley genome sequencing relied exclusively on shotgun sequencing of 88,731 BAC clones using high-throughput next-generation sequencing-by-synthesis22. This comprised 15,661 so-called gene-bearing BAC clones, preselected mainly by overgo-probe hybridization for the presence of transcribed genes and fingerprinted for definition of a minimum tiling path of the barley gene space. These gene-space minimum tiling path BAC clones were sequenced as combinatorial pools by Illumina short-read technology and, after quality trimming of de-convoluted reads, were assembled using Velvet version 1.2.09 as previously described13. The remaining 73,070 BACs were selected from a minimum tiling path representing the physical map of the barley genome16. Minimum tiling path BAC clones assigned to different barley chromosomes were sequenced at one of four sequencing centres, relying on highly multiplexed paired-end and mate-pair sequencing libraries using either the Roche 454 Titanium or the Illumina MiSeq, HiSeq2000 and HiSeq2500 platforms (Supplementary Note 1 and ref. 51). In brief, sequencing reads were de-convoluted on the basis of the used BAC-specific barcode sequence tags and assembled with sequencing centre-specific assembly pipelines. BAC clones sequenced on the Roche 454 Titanium platform were assembled with MIRA51 according to previously described procedures52,53. Illumina HiSeq2000 paired-end sequencing data (2 × 100 nucleotides) of BAC clones were assembled either with CLC Assembly Cell version 4.0.6 beta (http://www.clcbio.com/products/clc-assembly-cell/) set to default parameters12, SOAPdenovo version 2.01 (ref. 54) or the ABySS assembler (version 1.5.1)55. Sequence contigs of the de novo BAC assemblies larger than 500 base pairs (bp) were scaffolded using mate-pair sequencing information either generated from BAC DNA-derived 8 kbp insert mate-pair sequencing libraries or from 2 kbp, 5 kbp or 10 kbp genomic DNA-derived mate-pair libraries. This was achieved by either using BWA mem version 0.7.4 (ref. 56) with default parameters for read mapping, followed by scaffolding individual BACs using SSPACE version 3.0 Standard57, or with SOAPaligner/soap2 version 2.21 and using SOAPdenovo54 scaffolder version 2.01.

Genome-wide three-dimensional chromatin conformation capture sequencing

To generate physical scaffolding information for the BAC sequence based genome assembly, as proposed in ref. 21, Hi-C and tethered conformation capture (TCC) sequencing data were generated from 7-day-old leaf tissue of greenhouse-grown barley plantlets by adapting previously published procedures (Supplementary Note 2). In brief, for Hi-C, freshly harvested leaves were cut into 2 cm pieces and vacuum infiltrated in nuclei isolation buffer supplemented with 2% formaldehyde. Crosslinking was stopped by adding glycine and additional vacuum infiltration. Fixed tissue was frozen in liquid nitrogen and ground to powder before re-suspending in nuclei isolation buffer to obtain a suspension of nuclei. About 107 purified nuclei were digested with 400 units of HindIII as described previously58. Digested chromatin was marked by incubating with biotin-14-dCTP and Klenow enzyme using a fill-in reaction20 resulting in blunt-ended repaired DNA strands. Biotin-14-dCTP from non-ligated DNA ends was removed owing to the exonuclease activity of T4 DNA polymerase, followed by phenol–chloroform extraction and washing of the precipitated DNA as described20. As an alternative to Hi-C, the TCC protocol was also adapted for barley. Nuclei were prepared from barley leaf tissue as described above for Hi-C, before biotinylating the isolated chromatin using EZlink Iodoacetyl-PEG2-Biotin. The samples were neutralized with SDS, and DNA was digested with HindIII, dialysed, followed by immobilization to low surface coverage using streptavidin-coated magnetic beads19. Open DNA ends were labelled with biotin-14-dCTP using Klenow enzyme, and blunt-ended, labelled DNA products were collected from the magnetic beads by reversing the formaldehyde crosslink using proteinase K19. Biotin-14-dCTP from non-ligated DNA ends was removed by using Exonuclease III19. Hi-C and TCC products were mechanically sheared to fragment sizes of 200–300 bp by applying ultrasound using a Covaris S220 device followed by size-fractionation using AMPure XP beads. DNA fragments in the range between 150 and 300 bp were blunt-end repaired and A-tailed before purification through biotin–streptavidin-mediated pull-down58. Illumina paired-end adapters were ligated to the Hi-C and TCC products, respectively, followed by PCR amplification, pooling of PCR products and purification with AMPure XP beads before quantification of Hi-C/TCC libraries by qPCR for Illumina HiSeq2500 PE100 sequencing20.

Nanochannel-based genome mapping

Long-range scaffolding of genome sequence assemblies was facilitated by BioNano genome maps generated by nanochannel electrophoresis of fluorescently labelled high-molecular mass DNA obtained from flow-sorted chromosomes59. High-molecular mass DNA was prepared from 3.5 × 106 purified chromosomes (whole genome) of barley cultivar Morex essentially following published procedures60,61. The purified chromosomes were embedded in agarose miniplugs to achieve approximate concentrations of 1 million chromosomes per 40 μl volume before being treated with proteinase K as described previously61. DNA was labelled at Nt.BspQI nicking sites (GCTCTTC) by incorporation of fluorescent-dUTP nucleotide analogues using Taq polymerase as described previously59. The labelled DNA was analysed on the Irys platform (BioNano Genomics) in 191 cycles in total, generating 243 Gb of data exceeding 150 kb. On the basis of the label positions on single DNA molecules, de novo assembly was performed by a pairwise comparison of all single molecules and graph building62. The parameter set for large genomes was used for assembly with the IrysView software. A P value threshold of 10−9 was used during the pairwise assembly, 10−10 for extension and refinement steps and 10−14 for merging contigs. A whole-genome map of 4.3 Gb was obtained (Extended Data Table 1).

Data integration for constructing pseudomolecules

The construction of pseudomolecules representing the seven barley chromosomes followed an iterative, mainly automated procedure which involved the integration of the following major datasets: (1) sequence assemblies of 87,075 unique, successfully sequenced and assembled BAC clones; (2) BAC assembly information from a genome-wide physical map of barley16; (3) 571,814 end-sequences of BAC clones7; (4) a dense linkage map assigning genetic positions to 791,177 contigs of a whole-genome shotgun assembly of barley cultivar Morex17; (5) Hi-C/TCC sequence information; and (6) the optical map of the genome of barley cultivar Morex. A schematic outline of the procedure is presented elsewhere22. In the first step, overlaps between individual BAC assemblies were searched with Megablast63 by either applying ‘stringent’ or ‘permissive’ alignment criteria22 and by combining with the high density genetic map information. On the basis of this initial analysis, a BAC overlap graph was constructed by use of the R package igraph64 considering the above-listed additional datasets in subsequent iterative steps. Building the overlap graph focused first on overlaps obtained under ‘stringent’ search criteria for BACs within individual physical map contigs (FP contigs) and then subsequently also between independent FP contigs. Subsequently, overlaps obtained under ‘permissive’ criteria were evaluated while checking for cumulative evidences provided by the additional datasets supporting the overlap information22. Ordering and orienting of the resultant sequence scaffolds were achieved by integrating the overlap graph with Hi-C /TCC data22. Before the construction of pseudomolecules, we (1) identified genes incomplete or missing in the non-redundant sequence, but represented by (a) BAC sequence that had been excluded from the construction of the non-redundant sequence, or by (b) Morex WGS contigs, and (2) performed a final scan for contaminant sequences. Then a single FASTA file containing a single entry for each barley chromosome (a ‘pseudomolecule’) and an additional entry combining all sequences not anchored to chromosomes was constructed22.

Three-dimensional chromatin conformation analysis

Mapping of Hi-C/TCC reads and assignment to restriction fragments were performed as described elsewhere22. Briefly, raw reads were trimmed with cutadapt65. Trimmed Hi-C reads were mapped to the barley pseudomolecule sequence with BWA mem (version 0.7.12)66. Duplicate removal and sorting were performed with NovoSort (http://www.novocraft.com/products/novosort/). Mapped reads were assigned to restriction fragments with BEDtools67, tabulated with custom AWK scripts and imported into R (https://www.r-project.org/). Raw counts of Hi-C links were aggregated in 1 Mb bins and normalized separately for intra- and interchromosomal contacts using HiCNorm68. Contact probability matrices were plotted using standard R functions69. Principal component analysis was performed with the R function prcomp() on the matrix of log-transformed normalized Hi-C link counts between 1 Mb fragments.

We fitted the linear model log10(nl) ~ log10(dist) + abs(cen_dist1 – cen_dist2) + arm1:arm2 + apos1:apos1 using the R function lm(). Here, nl is the normalized link count between two 1 Mb bins, dist is their distance in the linear genome, cen_dist1 and cen_dist2 are the relative distances from the centromere of both loci, arm1 and arm2 are the chromosome arm assignment of both loci, and apos1 and apos2 are the relative distances of both loci from the ends of the chromosome arm (that is, apos1 is close to zero if locus 1 is either near the centromere or the telomere, and close to one if locus 1 resides in interstitial regions). TCC reads of Morex × Barke F1 hybrids were mapped to a synthetic reference representing the parental genomes. An in silico Barke assembly was created by inserting SNPs discovered by aligning Barke WGS reads to the Morex reference assembly with BWA MEM66 and calling variants with SAMtools70. SNPs were then inserted into the Morex reference using the FastaAlternateReferenceMaker of GATK71. TCC reads of the hybrid were then mapped to the synthetic reference as described above. Only uniquely alignable read pairs were considered. Hi-C link counts were tabulated at the level of chromosomes.

Fluorescence in situ hybridization was performed with H. vulgare nuclei as described earlier72 using Arabidopsis-type telomere and barley centromere-specific [AGGGAG]5 repeat probes73.

Automated annotation of transcribed regions

Automated gene annotation of the barley reference sequence assembly was based on four datasets providing independent gene evidence information (Supplementary Note 3). This included (1) RNA sequencing (RNA-seq) data; (2) reference protein predictions from barley7, rice74, B. distachyon75 and S. bicolor76; (3) published barley full-length complementary DNA (fl-cDNA) sequences77; and (4) newly generated barley PacBio Iso-Seq data. Previously published7 and newly generated RNA-seq datasets were derived from a total of 16 different tissues, each with three biological replicates, including seven vegetative, six inflorescence, two developing grain and one germinating grain tissues. RNA-seq libraries were sequenced on Illumina HiSeq2000 in paired-end 2 × 100 nucleotides (PE100) mode (Supplementary Note 3). To support gene calling in general, and the identification of alternative splice forms in particular, enriched full-length transcript information was generated by the Iso-Seq method using the PacBio RS II system and DNA Sequencing Chemistry 4.0 version 2 (Supplementary Note 3). RNA-seq-based transcript structures, reference-based gene model predictions, structure information from Iso-Seq alignments as well as structure information from flcDNA sequence alignments were clustered into a consensus transcript set using Cuffcompare78 (Supplementary Note 3). Predicted transcript sequences were automatically extracted into a single FASTA file on the basis of respective coordinates in the genome assembly. Putative open reading frames and corresponding peptide sequences, including prediction of Pfam domains, were obtained by applying TransDecoder (https://transdecoder.github.io), which also resulted in reports about predicted alternative peptides per transcript (Supplementary Note 3). A single best translation per transcript was selected on the basis of BLASTP79 comparison of all predicted peptides to a comprehensive protein database containing high-confidence protein sequences from A. thaliana80, maize47, B. distachyon75, rice75 and S. bicolor76, followed by additional filtering procedures (Supplementary Note 3). Functional descriptions (‘human readable descriptions’) were generated for all potential genes using the AHRD pipeline (https://github.com/groupschoof/AHRD) on the basis of one representative protein sequence for each gene locus. Gene candidates were then classified into high- and low-confidence genes and further subdivided into nine classes, each supported by different levels of gene evidence (Supplementary Note 3). High-confidence protein-coding genes either showed significant sequence homology to a reference protein or were associated with a predicted function. Low-confidence genes were characterized by (1) having no or only weak sequence homology to reference proteins and no predicted function, (2) they were candidates for transposons or (3) they lacked an open reading frame of a minimal length (Supplementary Note 3). Completeness of gene-space representation was evaluated with the BUSCO pipeline23 (Extended Data Fig. 2b).

Feature distributions along the chromosomes

A sliding window approach with a window size of 4 Mb and a shift of 0.8 Mb was used to display the distribution of different genome components and other features such as GC content or recombination rate along the chromosomes. The resulting data were smoothed with the python function scipy.signal.gaussian (p1 = 40, p2 = 10 for Fig. 1a; p1 = 15, p2 = 3 for Fig. 2a). The boundaries of genomic compartments (Fig. 1) are given in Supplementary Table 4.4.

Annotation of the non-genic part of the genome

Transposable elements were detected and classified by homology search with Vmatch (http://www.vmatch.de) against the REdat_9.7_Triticeae section of the PGSB transposon library81. The following parameter settings were used: identity ≥70%, minimal hit length 75 bp, seed length 12 bp (exact commandline: -d -p -l 75 -identity 70 -seedlength 12 -exdrop 5). The Vmatch output was filtered for redundant hits by prioritizing higher-scoring matches and then either shortening (<90% coverage and ≥50 bp rest length) or removing lower-scoring overlaps.

The identification of full-length LTR retrotransposons with LTRharvest82 resulted in 143,957 non-overlapping candidate sequences using the following parameter settings: ‘overlaps best -seed 30 -minlenltr 100 -maxlenltr 2000 -mindistltr 3000 -maxdistltr 25000 -similar 85 -mintsd 4 -maxtsd 20 -motif tgca -motifmis 1 -vic 60 -xdrop 5 -mat 2 -mis -2 -ins -3 -del -3’. All candidates were annotated for PfamA domains with hmmer3 software83 and stringently filtered for false positives by several criteria, the main ones being the presence of at least one typical retrotransposon domain (for example, RT, RH, INT, GAG) and a tandem repeat content below 25%. This resulted in a final set of 24,952 high-confidence full-length LTR retrotransposons. Insertion ages of the LTR retrotransposons were calculated according to the method of ref. 84 by the divergence of 5′ and 3′ LTRs that had been identical at the time of transposition. We used a grass-specific mutation rate of 1 × 10−8. The average age of all full-length LTR elements was calculated in 4 Mb windows and plotted in Fig. 1a. The frequencies of 20-mers were determined using Tallymer85.

Phylogenetic analysis of Gypsy elements was performed on predicted protein sequences deposited at the TREP database32. Protein domains in predicted open reading frames were identified with Pfam86, SignalP87 and COILS88.

For the analysis of transposable element content in up- and downstream regions of genes, 10 kb immediately flanking the predicted coding sequences of all high-confidence genes were extracted from the genome assembly. The genomic segments were then used in BLASTN searches79 against the TREP database32. After an initial annotation, previously unclassified or poorly characterized transposable element families were re-analysed and new consensus sequences were constructed. Analysis of up- and downstream regions was then repeated with the updated TREP database. The transposable element family producing the longest BLASTN hit was determined for every 20th base position of each 10 kb segment, resulting in 500 data points for each up- and downstream region of the high-confidence genes.

Gene family analysis

Gene family clusters were defined from 39,734 barley high-confidence class genes and the annotated gene sets of Rice MSU7.0 (39,049 genes, http://rice.plantbiology.msu.edu/), B. distachyon version 3.1 (31,694 genes, https://phytozome.jgi.doe.gov/pz/portal.html#!info?alias=Org_Bdistachyon), S. bicolor version 3.1 (33,032 genes, https://phytozome.jgi.doe.gov/pz/portal.html#!info?alias=Org_Sbicolor) and A. thaliana TAIR10 (27,416 genes, https://www.arabidopsis.org/)) using OrthoMCL89 software version 2.0. Splice variants were removed from the datasets, keeping only the representative/longest protein sequence prediction, and datasets were filtered for internal stop codons and incompatible reading frames. In the first step, pairwise sequence similarities between all input protein sequences were calculated using BLASTP79 with an e-value cut-off of 10−5. Markov clustering of the resulting similarity matrix was used to define the orthologue cluster structure, using an inflation value (-I) of 1.5 (OrthoMCL default). Gene families with barley-specific gene duplications, compared with other plant species, were extracted from the ENSEMBL Compara pipeline90. Over- and under-representation of Gene Ontology terms between barley and other plant species (Supplementary Tables 4.1–4.3) and between genomic compartments (Supplementary Table 4.5) were analysed with a hypergeometric test using the functions GOstats and GSEABase from the Bioconductor R package91 against a universe of all genes with Gene Ontology annotations. REViGO92, which removes redundant and similar terms from long Gene Ontology lists by semantic clustering, was applied to visualize the enrichment results. Expansion of three barley gene families encoding α-amylases, the vacuolar processing enzyme VPE2 protein subfamily and the sugar transporters SWEET11 subfamily, with specific importance in barley grain filling/seed development or barley germination/malting, were analysed in greater detail using BLAST searches (versus genome and gene prediction) as well as GenomeThreader mappings to the barley genome assembly. Further details are provided in Supplementary Note 4. In situ hybridizations for SWEET genes were performed as described previously93.

Analysis of sequence and haplotype diversity

Ninety-six two-row spring (n = 48) and winter (n = 48) homozygous inbred elite barley lines (Supplementary Table 5.1) were subjected to exome capture using the barley Roche NimbleGen exome capture liquid array94 and sequenced on the Illumina HiSeq 2500 platform. An average of 2 × 21,876,780 paired-end Illumina reads per sample was generated. This corresponds to approximately 72× coverage of the 61 Mb exome capture space.

The raw Illumina reads were mapped to the reference sequence with BWA-MEM version 0.7.10 (ref. 66), using a stringent mismatch setting of ≤2% mismatches per read. Variant calling was performed with the Genome Analysis Tool Kit (GATK)71 version 3.4.0, following the GATK Best Practices pipeline (https://www.broadinstitute.org/gatk/guide/best-practices.php). This included read de-duplication, indel realignment, base quality score recalibration and variant calling with the latest version of the HaplotypeCaller. The workflow was implemented in a BASH script. The Tablet assembly viewer95 was used for visual spot checks of mappings and SNPs calls.

Variant discovery resulted in 15,982,580 variants in total, of which 943,959 were multi-nucleotide polymorphisms or short insertions/deletions (indels), while the remainder represented SNPs. For subsequent genetic analysis, we first reduced the total variant dataset by applying rigorous filtering criteria to produce a highly robust subset of 72,563 SNPs distributed across all seven barley chromosomes. The filtering applied was as follows: (1) ≥8× coverage for ≥50% of the samples; (2) ≥95% of samples represented at each SNP locus; (3) ≥5% minor allele frequency at the level of the sample: that is, counting sample genotypes rather than individual reads; (4) a VCF SNP quality score ≥30; and (5) ≥98% of samples homozygous. These filters reduced false-positive variant calls by removing spurious variant calls resulting from systematic read mis-mapping. Of this filtered dataset, a subset of 3,500 randomly sampled markers from each chromosome was analysed with the Haploview software96. This subsampling was required as Haploview was unable to generate the required plots when larger data volumes were used as input. Haploview was run on defaults, using the GABRIEL blocks method. The genotype calls were also imported into the genotype visualization software Flapjack97 to produce chromosome-scale images of haplotype diversity within the spring and winter pools. Diversity statistics were calculated in GenAlEx version 6.502 (ref. 98) and rolling averages based on 100 adjacent SNPs were plotted in Microsoft Excel 2010.

Data availability

The genome assembly for barley has been deposited in the Plant Genomics and Phenomics Research Data Repository under digital object identifier http://dx.doi.org/10.5447/IPK/2016/34. Accession numbers for all deposited datasets are listed in Supplementary Note 1. The barley genome assembly has been deposited on the IPK Barley Blast Server (http://webblast.ipk-gatersleben.de/barley_ibsc/). All other data are available from the corresponding authors upon reasonable request.