De novo assembly, annotation, and comparative analysis of 26 diverse maize genomes

Abstract

We report de novo genome assemblies, transcriptomes, annotations, and methylomes for the 26 inbreds that serve as the founders for the maize nested association mapping population. The number of pan-genes in these diverse genomes exceeds 103,000, with approximately a third found across all genotypes. The results demonstrate that the ancient tetraploid character of maize continues to degrade by fractionation to the present day. Excellent contiguity over repeat arrays and complete annotation of centromeres revealed additional variation in major cytological landmarks. We show that combining structural variation with SNPs can improve the power of quantitative mapping studies. Finally, we document variation at the level of DNA methylation, and demonstrate that unmethylated regions are enriched for cis-regulatory elements that contribute to phenotypic variation.

One sentence summary:

A multi-genome analysis of maize reveals previously unknown variation in gene content, genome structure, and methylation.

Maize is the most widely planted crop in the world and an important model system for the study of gene function. The species is known for its extreme genetic diversity, which has allowed for broad adaptation throughout the tropics and intensive use in temperate regions. Nevertheless, most current genomic resources are referenced to a single inbred, B73, which contains only 63-74% of the genes and/or low-copy sequences in the full maize pan-genome (1–4). Moreover, there is extensive structural polymorphism in non-coding and regulatory genomic regions that has been shown to contribute to variation in numerous traits (5). In recent years, additional maize genomes have been assembled, allowing limited characterization of the species pan-genome (2, 6–10). However, comparisons across genome projects are often confounded by differences in assembly and annotation methods.

The maize Nested Association Mapping (NAM) population was developed to study the genetic architecture of quantitative traits (11). Twenty-five founder inbred lines were strategically selected from a larger association panel (12) to represent the breadth of maize diversity, including lines from the non-stiff-stalk temperate heterotic group, lines from tropical and subtropical regions of Africa, Asia, and the Americas, and both sweet corn and popcorn germplasm (13). Each NAM parental inbred was crossed to B73 and selfed to generate 25 populations of 200 recombinant inbred lines that combine the advantages of linkage and association mapping for important agronomic traits (14). Biological infrastructure continues to be developed around these lines (e.g. (15, 16)), but comprehensive genomic resources are needed to fully realize the power of the NAM population.

Consistency and quality of genome assemblies

Here we describe assembled and annotated genomes for the 25 NAM founder inbreds and an improved reference assembly of B73 (Table S1). The 26 genomes were sequenced to high depth (63-85X) using PacBio long-read technology, assembled into contigs using a hybrid approach (17), scaffolded using Bionano optical maps, and ordered into pseudomolecules using linkage data from the NAM recombinant inbred lines and maize pan-genome anchor markers (4). Assembly and annotation statistics improve upon nearly all available maize assemblies, including the previous B73 reference genome (18), with the total length of placed scaffolds (2.102-2.162 Gbp) at the estimated genome size of maize, a mean scaffold N50 of 119.2Mb (contig N50 of 25.7 Mbp), complete gene space (mean of 96% complete BUSCOs; (19)), and, based on the LTR Assembly Index (LAI, mean of 28; (20)), full assembly of the transposable-element-laden portions of the genome (Table 1; Table S2). Improvements in contiguity and completeness can be attributed to recent advances in sequence and optical map data, as well as more effective assembly algorithms (21).

Table 1:

Quality metrics for genome assemblies and gene model annotations. Darker shading indicates higher quality. The NAM lines are shaded based on their primary grouping (gold = stiff stalk heterotic group, blue = non-stiff-stalk heterotic group, gray = mixed tropical-temperate ancestry, purple = popcorn, orange = sweet corn, green = tropical).

^*Hp301 and P39 have the lowest amounts of TR-1 and subtelomere repeats, respectively. Our methods can overestimate assembly when repeats are in low abundance (17).

Gene identification and diversity in gene content

We sequenced mRNA from ten tissues for each inbred. These data were used for evidence-based gene annotation of each line, which was then improved using B73 full-length cDNA and expressed sequence tags (ESTs). The evidence set was augmented with ab initio gene models and the gene structures refined for all accessions using phylogeny-based methods. This pipeline revealed an average of 40,621 (standard error (SE) = 117) protein-coding and 4,998 (SE = 100) non-coding gene models per genome. The great majority of genes share orthologs with the grass (Poaceae) family and species in the Andropogoneae tribe of grasses, which includes maize and sorghum (Fig. 1A). The accuracy of the annotations, measured by the congruence between annotations and supporting evidence (Annotation Edit Distance, AED) (22), is higher than previous reference maize annotations (Fig. S1) (2, 6, 10, 18, 23).

Figure 1.

Pan genome analysis of the gene space. A) Pan-genes categorized by annotation method and phylostrata. Genes annotated with evidence have mRNA support whereas ab initio genes are predicted based on DNA sequence alone. Genes within progressing phylostrata - species Zea mays (maize), tribe Andropogoneae, family Poaceae, kingdom Viridiplantae - are more conserved. B) Number of pan-genes added with each additional genome assembly. Order of genomes being added into the pan genome was bootstrapped 1000 times. Tropical lines include (CML52, CML69, CML103, CML228, CML247, CML277, CML322, CML333, Ki3, Ki11, NC350, NC358, Tzi8), temperate lines include (B73, B97, Ky21, M162W, Ms71,Oh43, Oh7B, HP301, P39, and Il14H). C) Proportion of pan-genes in the core, near core, dispensable, and private fractions of the pan-genome. For B and C, tandem duplicates were considered as a single pan-gene and coordinates were filled in when a gene was not annotated, but an alignment with greater than 90% coverage and 90% identity was present within the correct homologous block. D) Number of tissues with expression (RPKM>1) for each gene in each genome based on their pan-genome classification. Tissues in this analysis include (root, shoot, V11 base, V11 middle, V11 tip, anther, tassel, and ear).

We next assessed the gene catalog of the pan-genome. Genes with high sequence similarity, located within blocks of homologous sequence in pairwise comparisons, were grouped together as one pan-gene. In many instances, a gene was not annotated by our computational pipeline, yet at least 90% of the gene was present in the correct homologous location; when this occurred, the pan-gene was considered present (Fig. S2 A; (17)), even though in some cases the absence of annotation may reflect fractionation and/or pseudogenization.

Across the 26 genomes, a total of 103,033 pan-genes were identified. Previous analysis reported ~63,000 pan-genes based on transcriptome assemblies of seedling RNA-seq reads from 500 individuals (1). The superior contiguity of our assemblies and the application of both ab initio and evidence-based annotation using RNA-seq from a diverse set of ten tissues, likely accounts for the increased sensitivity. Over 80% of pan-genes were identified within just ten inbred lines based on a bootstrap resampling of genomes (Fig. 1B). When considered separately, temperate and tropical lines have differentiated sets of pan-genes but show a comparable rate of pan-gene increase as lines are added, suggesting they have similar gene-content diversity (Fig. 1B).

Pan-genes, excluding tandem duplicates (17), were classified as core (present in all 26 lines), near-core (present in 24-25 lines), dispensable (present in 2-23 lines), and private (present in only one line) (Fig. 1C). The portion of genes classified into each of these groups was consistent across genotypes, with an average of 58.41% (SE = 0.07%) belonging to the core genome, 8.23% (SE = 0.05%) to the near-core genome, 31.75% (SE = 0.09%) to the dispensable genome, and 1.60% (SE = 0.08%) private genes (Fig. 1C; Fig. S2 B–C; Table S3). In total, there are 32,052 genes in the core/near-core portion of the pan-genome and 70,981 genes in the dispensable/private portion. The core genes (and gene families enriched for core genes (Table S4)) are generally from higher phylostrata levels (i.e. Viridiplanteae and Poaceae), while those in the near-core and dispensable sets either share orthologs only with closely related species or are maize-specific (Fig. S2 E). Some private genes may be spurious annotations resulting from imperfect masking of repeat sequences, as the majority of core/near-core genes are syntenic to sorghum (57.78%), whereas this is rarely the case for dispensable/private genes (1.83% syntenic). Core genes were expressed in more tissues (Fig. 1D) and had higher transcript abundance (Fig. S2F) when compared to genes present in fewer individuals. However, across the relatively small number of tissues (≥ 8 per line) profiled for this analysis, 18% of dispensable and 32% of private genes were expressed in at least one tissue. A total of 16,751 pan-genes were tandemly duplicated in at least one genome, of which 7,040 were duplicated in a single genome. On a per gene basis in genomes with at least one tandem duplicate the average copy number is 2.20 (SE = 0.01) (Fig. S2 D).

Partial tetraploidy and tempo of fractionation

The maize ancestor underwent a whole-genome duplication (WGD) allopolyploidy event 5-20 MYA ((24, 25), Fig. 2A). Evidence for WGD is found in the existence of two separate genomes that are broken and rearranged, yet still show clear synteny to sorghum (24, 26). Many duplicated genes have since undergone loss, or fractionation, reducing maize to its current diploid state (26, 27). Further, fractionation is biased towards one homoeologous genome (M2, more fractionated) over the other (M1, less fractionated) (26). The M1 and M2 subgenomes are composed almost exclusively of core (87.25%) and near-core (6.19%) pan-genes (Figs. 1C, 2A). The broad architecture of syntenic regions relative to sorghum is consistent across the NAM genomes (Fig. S3).

Figure 2.

The tempo of fractionation in maize. A) Schematic showing how genes were categorized. 16,195 conservatively chosen orthologs were subdivided into classes representing retained pairs, ancient fractionation, and recent fractionation. B) Unfolded site frequency spectrum (SFS) of segregating exon loss and non-coding SNPs (genic and non-genic) using sorghum to define the ancestral state. C) Heatmap of the number of co-retained exons between any two NAM lines. Lines with mixed ancestry (M37W, Mo18W, Tx303) are excluded. Colors indicate the Z-score (the difference measured in standard deviations between a single pairwise comparison and all others in the row).

Given the ancient timeframe of the WGD in maize and the rapid tempo of fractionation observed in other species (28, 29), little variation in the retention of specific homoeologs is expected at the species level. In fact, prior work in temperate maize suggested that most fractionation occurred before domestication (6, 30). However, our diverse set of genomes allows for a more complete characterization of fractionation within the species. Since fractionation can occur at the level of small deletions (27, 31), we evaluated both partial and complete homoeolog loss beginning with a conservative set of 16,195 maize pan-orthologs. We determined that 7,043 were single-copy orthologs, where the homoeologous gene was likely deleted prior to maize speciation (Fig. 2A). In addition, we identified 4,576 homoeologous pairs (Fig. 2A) of which 2,155 had the same exon structure of the sorghum ortholog in both homoeologs. In 1,281 pairs, at least one copy of the gene differed from its sorghum ortholog, but did not vary among NAM lines, likely representing fractionation that pre-dated Zea mays. Another 1,140 pairs varied across the genomes in their pattern of exon retention, segregating for deletions or structural differences in at least one copy of the gene. This segregating set was manually curated (Dataset S1) to remove loci where exons or flanking sequence could not be confidently identified (Fig. 2A), resulting in a curated set of 494 homoeolog pairs segregating for fractionation, which represents more than 10% of pairs present in the pan-genome. Of these, 281 M2 homoeologs had exon loss compared to 236 M1 homoeologs, a 19% difference (p < 0.05, χ² test), suggesting ongoing biased fractionation. Analysis of gene ontology terms revealed putative functional differences between fully fractionated and segregating fractionated loci (Fig. S4, Dataset S1).

Population genetic theory predicts mutations segregating within a species, like the segregating fractionation deletions we have identified, arose within the last 4N_e generations, where N_e represents the effective population size of the species. Using the N_e of the maize progenitor teosinte as an upward bound for maize (N_e = 150,000; (32)), we can infer that the majority of segregating fractionation arose within the last 600,000 generations. Therefore, the majority of segregating fractionation substantially post-dates the WGD. Theory also predicts that rare deletions should be younger than those segregating at intermediate frequency. We constructed the unfolded site frequency spectrum (SFS) of segregating fractionation deletions and compared this to the unfolded SFS of non-coding SNPs using sorghum to define the ancestral state (Fig. 2B). The data reveal a similar frequency distribution in deletions and SNPs, with a preponderance of rare variants in both, suggesting that a subset of fractionation may be quite young, with diploidization potentially continuing in modern maize. We also evaluated patterns of co-exon-retention in non-stiff-stalk temperate, tropical, and flint-derived maize, observing population-specific fractionation (Fig. 2C). This variation in homoeolog retention at the population level confirms previous suppositions about the tempo of fractionation (33) and may reflect relaxed constraint on retained homoeologs following domestication and migration of maize to temperate climates.

The repetitive fraction of the pan-genome

Transposable elements (TEs) were annotated in each assembly using structural features and sequence homology (34). Individual TE libraries from each inbred were then combined to form a pan-genome library, which was used to identify TE sequences missed by individual libraries. The annotations reveal that DNA transposons and LTR retrotransposons comprise 8.5% and 74.4% of the genome, respectively (Table S5, Fig. S5). A total of 27,228 TE families were included in the pan-genome TE library, of which 59.7% were present in all 26 NAM founders and 2.5% were unique to one genome (Fig. S6). The average percentage of intact and fragmented TEs were 30.5% and 69.5% (SE = 0.06%), respectively. As reported previously, Gypsy LTR retrotransposon families are more abundant in pericentromeric regions, while Copia LTR retrotransposons are enriched in the gene-dense chromosome arms (Fig. S7) (35). Tropical lines have significantly more Gypsy elements than temperate lines (p = 0.002, t-test), with mean Gypsy content of 1,018 Mbp and 988 Mbp, respectively (Table S5, Fig. S5). This may reflect increasing constraint on Gypsy proliferation in temperate lines that have, on average, smaller genomes (Table 1).

In some maize lines, over 15% of the genome is composed of tandem repeat arrays including the centromere repeat CentC, the two knob repeats knob180 and TR-1, subtelomere, and telomere repeats (36, 37). Repeats of this type remain a major impediment to assembly. A mean of 60% of CentC, 70% of the 4-12-1 subtelomeric sequence (38)), 28.9% of TR-1, 1% of knob180, and 0.09% of rDNA repeat units were incorporated in the final assemblies (Table 1).

A total of 110 (of 260) functional centromeres identified by CENH3 ChIP-seq (39, 40) were fully assembled, and of these 88 are gapless ((Fig. S8A and (40)). Chromosomes with very long CentC arrays (such as chromosomes 1, 6, and 7) often have assembly gaps and the precise location of the centromere could not be determined. However, many centromeres either have fully assembled small CentC arrays or the functional centromeres are located to one side of the CentC tracts in regions dominated by retrotransposons (Fig. 3A). By projecting all centromere locations onto B73, we were able to identify twelve centromere movement events (three on chr5 and chr9, and two on chr3, chr8 and chr10), clarifying and extending prior evidence for centromere shifting (39) (Fig. 3B, Fig. S8B). The variation in CentC abundance and positional polymorphism made it possible to gaplessly assemble at least two variants of all ten centromeres (Fig. S8A).

Figure 3.

Structural variation in the NAM founders. A) Pairwise alignments between Ki11, B73, Il14H on chromosome 8. Grey links represent syntenic aligned regions; gaps of unknown size (scaffold gaps) are marked by dashed lines. B) Large (>100 kbp) structural variants, centromeres, and knobs across the NAM lines versus the B73 reference. The subset of SVs larger than 1 Mbp were manually curated, and only those containing genes are represented. Features 1-5 highlight major SVs: 1) Multiple centromere movement events; 2) A major inversion previously hypothesized based on suppressed recombination; 3) A large deletion in the Ms71 inbred; 4) Knob polymorphism; 5) Reciprocal translocation between chromosome 9 and 10 in the Oh7B inbred (both segments placed in their standard positions for display).

Both knob180 and TR-1 arrays are subject to meiotic drive and accumulate when a chromosome variant known as Abnormal chromosome 10 (Ab10) is present (37, 41). Although Ab10 is absent from modern inbreds, its legacy remains in the form of many large knobs. The majority of knob180 and TR-1 repeat arrays were identified in mid-arm positions (81.9%) where meiotic drive is most effective. Long knob180 and TR-1 repeat arrays can occur separately, but are more frequently intermingled in fragmented arrays along with transposons (Fig. 3A, Fig. S9) (42). Analysis of classical (cytologically visible) knobs on chromosome 1S, 2S, 2L, 3L, 4L, 5L, 6L, 7L, 8L, and 9S revealed that their locations are syntenic and that several are composed of a series of disjointed smaller knobs (Fig. 3A, Fig. S10). In some lines, knobs are not visible cytologically but can still be detected as smaller arrays at the sequence level; however, many show strict presence-absence variation among the NAM founder inbreds.

Tandem repeat arrays are also commonly found at the ends of chromosome arms (Table S6). Among the 520 chromosome ends, 57.9% contained knob180 repeats and 30.5% contained subtelomere repeats. At least 65.6% of chromosome ends were fully assembled as indicated by the presence of telomere sequences.

Structural variation and impact on phenotype

Comparative analyses among the NAM genotypes to B73 revealed a cumulative total of 791,101 structural variants (SVs) greater than 100 bp in size. Tropical lines, which are the most divergent from B73, include a substantially higher number of SVs than temperate lines (mean = 32,976 versus 29,742; p = 0.00013) (Tables S7, S8). Structural variants are more common on chromosome arms where recombination is highest (Fig. S11), similar to SNPs and other forms of genetic variation (43). Almost half (49.6%) of SVs were <5 kbp in size, with 25.7% being less than 500 bp. Across all size classes SVs are skewed toward rare variants (Fig. S12). Several large SVs were found segregating within the 26 NAM genomes (Fig. 3B), including 35 distinct inversion polymorphisms and 5 insertion-deletion polymorphisms >1 Mbp. For example, a 14.6 Mbp inversion on chromosome 5 in the CML52 and CML322 lines, which was previously hypothesized based on suppressed recombination in the NAM RILs (11), is confirmed here based on assembly. Additionally, there is a 1.9 Mbp deletion with seven genes on chromosome 2 in the MS71 inbred, and a 1.8 Mbp deletion with two genes on chromosome 8 found in eight lines. Our data also capture a very large reciprocal translocation (involving >47 Mbp of DNA) between the short arms of chromosomes 9 and 10 in Oh7B that had been previously detected in cytological studies (38) (Fig. 3B).

The high proportion of rare SVs in maize suggests these may be a particularly deleterious class of variants, as observed in other species (44, 45). Indels and inversions occur in regions that have 49.8% fewer genic base pairs than the genomic background. Furthermore, SVs are 17% less likely to be found in conserved regions than SNPs (odds ratios of 0.27 and 0.58 for SVs and SNPs, respectively, Fisher’s Exact Test, p < 0.001). Approximate Bayesian computation modeling revealed that selection against SVs is at least as strong as that against nonsynonymous substitutions (Fig. S13; See Supplemental Methods). These results suggest that, when they occur, SVs are particularly consequential and relevant to fitness.

To estimate the phenotypic impact of SVs, we assessed the genetic basis of 36 complex traits (14) using 71,196 filtered SVs in 4,027 recombinant inbred lines derived from the NAM founder inbreds (11) (Fig. S14A). The analysis revealed that SVs explain a high percentage of phenotypic variance for disease traits (60.10% ~ 61.75%) and less for agronomic/morphological (20.04% ~ 61.04%) and metabolic traits (4.79% ~ 26.78%). Much of the phenotypic variation was also explained by SNPs, which were much more numerous (288-fold more) relative to our conservative set of SVs (Fig. S14A). When the SNP and SV data were integrated into one linear mixed model, the combined markers only slightly surpassed values from SNPs, consistent with the fact that most SVs are in high linkage disequilibrium with SNPs (Fig. S14A).

We also carried out genome-wide association analyses (GWAS) to identify specific SVs contributing to phenotypic variation for the same suite of traits (Fig. S14B–G). Among the detected GWAS signals, 93.05% overlapped with those identified with SNPs and 6.95% were unique to SVs (no significant SNP detected within 5 Mbp of significant SVs). There was a significant enrichment of SVs associated with phenotypes in genic regions (z = 8.022, p < 1.04e-15; Fig. S15). The most significant association between a SV and a trait not identified using SNP markers was a QTL for northern leaf blight (NLB) on chromosome 10 (Fig. S14F). This SV is within a gene encoding a thylakoid lumenal protein; such proteins could be linked to plant immunity through the regulation of cell death during viral infection (46). We anticipate that the effects of SVs may be even more pronounced in larger association panels where extensive historical recombination may help disentangle their effects from nearby SNPs.

Disease resistance in plants is frequently associated with SV in the form of tandem arrays of resistance genes. Complex arrays of resistance genes are retained, potentially through birth-death dynamics in an evolutionary arms race with pathogens, or through balancing selection for the maintenance of diverse plant defenses (47). Nucleotide-binding, leucine-rich-repeat (NLR) proteins provide a common type of resistance. Our data reveal that there are fewer NLR genes in maize than other Poaceae (Fig. S16) and that most NAM lines have lost the same clades of NLRs as sorghum (Fig. S17). Only one line (CML277) retains the MIC1 NLR clade, which is particularly fast-evolving in Poaceae (48). Nevertheless, there is clear NLR variation among the NAM lines (Fig. S18), and tropical genomes contain a significantly higher number of NLR genes than temperate genomes (t-test, p=0.006), suggesting ongoing co-evolution with pathogens, particularly where disease pressure is high.

The annotated NLR genes were significantly enriched for overlap with SVs (boot-strap permutation test, p<0.001). An extreme example is found at the rp1 (resistance to Puccinia sorghi1) locus on the short arm of chromosome 10, which is known to be highly variable (49). We observed exceptional diversity in the NAM lines with as few as 4 rp1 copies in P39, and as many as 30 in M37W (Table S9). However, due to its repetitive nature, only 18 NAM lines have gapless assemblies of the rp1 locus.

SVs linked to transposons have been shown, through the modulation of gene expression, to underlie flowering-time adaptation in maize during tropical-to-temperate migration (50, 51). Our SV and TE-annotation pipelines identified the adaptive CACTA-like insertion previously reported upstream of the flowering-time locus ZmCCT10 (51). We also surveyed 173 genes linked to flowering-time (52, 53) and discovered three genes (GL15, ZCN10, and Dof21) with TE-derived SVs <5 kbp upstream of their transcription start sites. These SVs distinguish temperate from tropical lines (t < −2.346, p < 0.0358) (Fig. S19) and show significant correlation (F > 8.658, p < 0.001) with expression levels.

Discovery of candidate cis-regulatory elements through DNA methylation

Based on sequence alone, it can be difficult to identify functional sequences in the intergenic spaces. One approach is to score for unmethylated DNA, which provides both a tissue-independent indicator of gene regulatory elements and evidence that annotated genes are active (5, 54, 55). We sequenced enzymatic methyl-seq (EM-seq) libraries from each NAM line and identified methylated bases in three sequence contexts, CG, CHG, and CHH (where H = A, T, or C). Results are consistent across genes and transposons, demonstrating the quality of the libraries (Figs. S20, S21). There is minor variation in total methylation across inbreds, with CML247 being noteworthy for uniformly lower CG methylation in several tissues (Fig. S22). Such natural variation in methylation is also observed in Arabidopsis ecotypes (56).

Each of the three methylation contexts reveal information on the locations of repeats, genes and regulatory elements. mCHH levels are generally low except at heterochromatin borders, whereas mCHG and mCG are abundant in repetitive regions. Both mCHG and mCG are depleted from regulatory elements and mCHG is depleted from exons (57). However mCG is often present in exons (Fig. 4) (58). Thus, to identify unmethylated regions (UMRs) corresponding to regulatory elements and gene bodies, we defined UMRs using a method that takes into account mCHG and mCG but does not exclude high mCG-only regions (the term UMR is used for simplicity; some regions contain CG methylation). Comparison of the 26 methylomes revealed uniformity in number and length of UMRs, averaging about 180 Mbp in total length in each genome (Figs. S23, S24). To confirm the accuracy of the UMR data, we also identified accessible chromatin regions (ACRs) using ATAC-seq for each inbred. We expect chromatin to be accessible mainly in the subset of genes expressed in the tissue sampled (primarily leaves) and to show concordance with UMRs. The data reveal that a mean of 99% of genic and 96% of non-genic (distal) ACRs overlap with UMRs in each genome (Figs. S25, S26).

Figure 4.

UMR variation across the NAM founders. A) Annotation of the Miniature seed1 gene in the Mo17W inbred. An image from the MaizeGDB browser shows gene, TE, and UMR tracks. TE tracks are color-coded by superfamily: green/grey = LTR, red = TIR, blue = LINE. The grey vertical lines show 2.5 kbp intervals. B) Annotation and underlying methylation data for Miniature seed1 in the B73 inbred. The insertion of a Gypsy element moved part of the proximal UMR to a position 14 kbp upstream from the transcription start site (TSS). Methylation tracks indicate base-pair level methylation values from 0 to 100%. Asterisks indicate gaps in coverage, which are visible in separate tracks (Fig. S28). C) Relationship between methylation and gene expression. UMRs were mapped to B73 to identify UMRs that overlap with TSS. The Y axis indicates the ratio of transcripts per million (TPM, compared to B73) when the region is methylated (red) or unmethylated (teal).

To assess methylation diversity, we mapped UMRs from all inbreds to the B73 genome. Approximately 95% of genic UMRs overlap across genomes in pairwise comparisons (Fig. S27). UMR polymorphism is higher in the intergenic space, particularly among UMRs greater than 5 kbp from genes, where typically ~75% of UMRs overlap (Fig. S27). Even when the UMR sequence is conserved, its position relative to the closest gene may vary dramatically among inbreds. This is exemplified by the Miniature Seed1 gene where a UMR proximal to the promoter in Mo18W is displaced nearly 14 kbp upstream in B73 by a single Huck element (Gypsy LTR superfamily) (Fig. 4). The Huck insertion is present in 23 of 26 genomes, and in two of these (Oh43 and CML322), additional nested TE insertions increased the distance between the gene and the UMR to 27 kbp. Although UMR polymorphism correlates with genetic distance across NAM lines (Fig. S29), UMRs from Tzi8 were not substantially shared with other tropical genomes.

Adaptive variation in DNA methylation has been observed in maize (59), most likely through effects on gene expression. To estimate how well UMRs predict transcription, we identified a conservative subset of UMRs overlapping genes that were unmethylated in B73 but methylated in at least one other methylome. These differentially methylated regions were strongly correlated with differences in gene expression (Fig. 4, Fig. S30). We further evaluated the enrichment of significant GWAS SNPs across 36 traits in UMRs. Based on genome-wide estimates, UMRs show 2.50- to 3.26-fold enrichment across traits for significant associations. Roughly 18% of SNPs identified by GWAS lie outside of genic regions but within UMRs (Table S10), consistent with the view that UMRs can be used to identify functional, non-coding regions (5, 54, 55).

Summary

Our analysis of 26 genomes uncovered variation in both the genic and repetitive fractions of the pan-genome. Tropical, temperate, and flint-derived popcorn/sweet corn germplasm are differentiated in a number of striking ways including their pan-gene complement, homoeolog retention post-polyploidy, abundance of transposable elements, NLR disease-resistance gene copy number, and methylation profiles. The available data will have broad utility for genetic and genomic studies and facilitate rapid associations to phenotyping information. For example, the genic presence-absence variation identified here may be imputed across additional mapping populations to clarify its contribution to heterosis through complementation (60). More generally, these resources should motivate a shift away from the single reference mindset to a multi-reference view where any one of 26 inbreds, each with different experimental and agronomic advantages, can be deployed for the purposes of basic discovery and crop improvement.

Data and materials availability:

Genome assemblies and annotations can be accessed at https://maizegdb.org/NAM_project. Raw data used for the assemblies including PacBio, Illumina, and Bionano data are available through ENA BioProject IDs PRJEB31061 and PRJEB32225. RNA-Seq data is available at ENA ArrayExpress E-MTAB-8633 and E-MTAB-8628. EM-Seq reads are available at ENA ArrayExpress E-MTAB-10088. ATAC-seq reads are available under NCBI GEO accession GSE165787. Other files, tables and supplemental data can be found in CyVerse https://datacommons.cyverse.org/browse/iplant/home/shared/NAM/NAM_genome_and_annotation_Jan2021_release. Links to the NLR trees can be found at https://itol.embl.de/shared/xCJbI9ndshEK. Scripts used to generate and analyze data are available as a Zenodo repository (61).