Polyploidy or whole-genome duplication provides genomic opportunities for evolutionary innovations in many animal groups and all flowering plants1,2,3,4,5, including most important crops such as wheat, cotton and canola or oilseed rape6,7,8. The common occurrence of polyploidy may suggest its advantage and potential for selection and adaptation2,3,9, through rapid genetic and genomic changes as observed in newly formed Brassica napus10, Tragopogon miscellus11 and polyploid wheat12, and/or largely epigenetic modifications as in Arabidopsis and cotton polyploids5,13. Cotton is a powerful model for revealing genomic insights into polyploidy3, providing a phylogenetically defined framework of polyploidization (~1.5 million years ago (Ma))14, followed by natural diversification and crop domestication15. The evolutionary history of the polyploid cotton clade is longer than that of some other allopolyploids, such as hexaploid wheat (~8,000 years)12, tetraploid canola (~7,500 years)16 and tetraploid Tragopogon (~150 years)11. Polyploidization between an A-genome African species (Gossypium arboreum (Ga)-like) and a D-genome American species (G. raimondii (Gr)-like) in the New World created a new allotetraploid or amphidiploid (AD-genome) cotton clade (Fig. 1a)14, which has diversified into five polyploid lineages, G. hirsutum (Gh) (AD)1, G. barbadense (Gb) (AD)2, G. tomentosum (Gt) (AD)3, G. mustelinum (Gm) (AD)4 and G. darwinii (Gd) (AD)5. G. ekmanianum and G. stephensii are recently characterized and closely related to Gh17. Gh and Gb were separately domesticated from perennial shrubs to become annualized Upland and Pima cottons15. To date, global cotton production provides income for ~100 million families across ~150 countries, with an annual economic impact of ~US$500 billion worldwide6. However, cotton supply is reduced due to aridification, climate change and pest emergence. Future improvements in cotton and sustainability will involve use of the genomic resources and gene-editing tools becoming available in many crops9,18,19.

Fig. 1: Sequencing features of five cotton allotetraploid species.
figure 1

a, Evolution and domestication of five polyploid lineages, Gh, Gb, Gt, Gd and Gm, after polyploidization between an A-genome African species (Ga-like) and a D-genome American species (Gr-like). Typical seeds from each species are shown. The divergence time estimates are based on 21,567 single orthologs among the 5 species by using the synonymous substitution rate (r) of 3.48 × 10−9 (Methods and Supplementary Note). Scale bar, 10 mm; ka, thousand years ago. b, Chromosomal features and synteny of the Gm genome. Notes in circos plots: (i) estimated lengths of 13 A and 13 D homoeologous pseudochromosomes; (ii) distribution of annotated genes; (iii) TE content (Gypsy, steel blue; Copia, grey; other repeats, orange); (iv,v) stacked SNP (iv) and indel (v) densities between Gm and Gb, Gd, Gh and Gt, respectively (see inset), and (vi) syntenic blocks between the homoeologous A and D chromosomes. The densities in plots in (ii)–(v) are represented in 1 Mb with overlapping 200-kb sliding windows. c, Genome-wide syntenic relationships among A and D subgenomes in five allotetraploids relative to the A-genome-like Ga (A2 genome) and D-genome-like Gr (D5 genome). Structural variations among syntenic blocks are marked with colored ribbons.

Source data

Cotton genomes have been sequenced for the D-genome (Gr)20 and A-genome (Ga)21 diploids and two cultivated tetraploids22,23,24,25,26. These analyses have shown structural, genetic and gene expression variation related to fiber traits and stress responses in cultivated cottons, but the impact of polyploidy on selection and domestication among the wild and cultivated polyploid cotton species remains poorly understood6. Here we report high-quality genomes for all five allotetraploid species and show that despite wide geographic distribution and diversification, allotetraploid cotton genomes retained the syntenic gene content and genomic diversity relative to respective extant diploids. Evolutionary rate heterogeneities, gene loss and positively selected genes characterize the two subgenomes of each species but differ among polyploid lineages. Transposable elements (TEs) are dynamically exchanged between the two subgenomes, facilitating genome-size equilibration following allopolyploidy. Gene expression diversity in the fiber tissues involves selection, coexpression networks and N6-methyladenosine (m6A) RNA modifications. In cultivated polyploid cottons, recombination suppression correlates with DNA hypermethylation and weak chromatin interactions and can be overcome by wild introgression and possibly epigenetic remodeling. The results offer unique insights into polyploid genome evolution and provide valuable genomic resources for cotton research and improvement.


Sequencing, assembly and annotation

Sequencing of the five allotetraploid cotton genomes entailed using complementary whole-genome shotgun strategies, including sequencing by single-molecule real-time (PacBio SEQUEL and RSII, ~440× genome equivalent), Illumina (HiSeq and NovaSeq, ~286×) (Supplementary Dataset 1a) and chromatin conformation capture (Hi-C seq) (~326×) (Methods). Homozygous single nucleotide polymorphisms (SNPs) and insertions/deletions (indels) were also used to correct the consensus sequence (Supplementary Dataset 1b,c). The rate of anchored scaffolds is 97% in Gb and 99% or higher in the other 4 species. Scaffolds were oriented, ordered and assembled into 26 pseudo-chromosomes with very low (0.1–0.8%) gaps (Table 1 and Supplementary Dataset 1d). The assembled genomes range in size from 2.2 to 2.3 gigabase pairs (Gbp; Table 1), slightly smaller than the sum of the two A- and D-genome diploids (1.7/A + 0.8/D ≈ 2.5 Gbp/AD)20,21. Nearly 73% of the assembled genomes are repeats and TEs (Supplementary Dataset 1e), predominantly in pericentromeric regions in Gm (Fig. 1b) and the other 4 species (Extended Data Fig. 1). The completeness and contiguity of these genomes compare favorably with Sanger-based sequences of sorghum27 and Brachypodium28.

Table 1 Genome assembly and annotation statistics for five allotetraploid cotton species

The euchromatic sequences of 5 polyploid genomes are complete (Supplementary Note), as supported by BUSCO scores (>97%) and 36,880 (>99%) primary transcripts from the Gr version 2 release20 (Supplementary Dataset 1b), with the number of protein-coding genes predicted to range from 74,561 (Gb) to 78,338 (Gt; Table 1), which are 3,000–4,000 more than reported in Gh and Gb23. Although the A subgenome (1.7 Gbp) is twice the size of the D subgenome (0.8 Gbp)20,21, mirroring the ancestral state of their extant diploids, the two have similar numbers of protein-coding genes (ratio of D/A ≈ 1.06; Supplementary Dataset 1f).

As an indication of the improved contiguity (Supplementary Note), the contig length in the Gh genome increases 6.9-fold with a 7.7-fold reduction in fragmentation (6,733 versus 51,849), compared to the published sequences22. The improvement is substantial in the Gb genome with a 15.9-fold reduction in N50 contigs and a 23-fold increase in N50 contig length (from 77.6 to 1,800 kilobase pairs (kb)). Moreover, most quality scores are 2-5-fold higher in the 3 wild polyploid species than in Gh and Gb (Table 1).

Reciprocal 24-nucleotide masking and syntenic analyses show that our Gh and Gb assemblies have ~23- and 2.7-fold more unique sequences, respectively, than the published ones22 also with variable gap sizes (10–200 kb; Extended Data Fig. 2a). Some specific genes are present in our annotations and the published data, which are largely related to gene copy number variation (more decreases than increases). Other differences include inversions (132–133 megabase pairs (Mb)) with two large ones (A06 and D03) present in similar regions of both Gh and Gb22 (Extended Data Fig. 2b), which could result from errors and/or unresolved alternative haplotypes; these inversions were confirmed using Hi-C data (Extended Data Fig. 2c). Notably, the published Hai7124 strain22 is a Gb local strain that is different from Gb 3-79, and Gh TM-1 strains may vary; these can also contribute to the observed variation.

Evolution within and between five polyploids

Using the diploid20,21 and 5 polyploid cotton genomes, we estimated divergence at 58–59 Ma between Gossypium and its relative Theobroma cacao (Extended Data Fig. 3a and Supplementary Note), 4.7–5.2 Ma between the extant diploids (Extended Data Fig. 3b), and 1.0–1.6 Ma between polyploid and diploid clades. Genome-wide phylogenetic analysis (Extended Data Fig. 4a) supports a monophyletic origin for the five allotetraploid species29. Within the polyploid clade, the highest divergence (~0.63 Ma) occurs between Gm and the other 4 species, with the most recent divergence (~0.20 Ma) between Gb and Gd. This genomic diversification was accompanied by biogeographic radiation to the Galapagos Islands (Gd), the Hawaiian Islands (Gt), South America (northeastern Brazil) (Gm)30, Central and South America, the Caribbean, and the Pacific (Gh and Gb)31, with separate distribution and domestication of diploid cultivated cottons in southern Arabia, North Africa, western India and China32 (Extended Data Fig. 4b). Over the last 8,000 years, Upland (Gh) and Pima (Gb) cottons were independently domesticated in northwest South America and the Yucatan Peninsula of Mexico, respectively, under strong human selection, leading to the modern annualized crops15.

After whole-genome duplication, duplicate genes may be lost or diverge in functions33, but the pace of this process has rarely been studied in allopolyploids. Using 17,136 homoeolog pairs shared among all 5 allotetraploid species, we demonstrate that most (14,583, 85.5%) homoeolog pairs evolved at statistically indistinguishable rates throughout the polyploid clade relative to the diploids (Supplementary Dataset 2a), but those with rate shifts occur more commonly in the A (1,476, 8.5%) than in the D (845, 5%) subgenome. We further revealed that the D homoeologs generally acquire substitution mutations more quickly than the A homoeologs in most lineages, whereas the Gh and Gt lineages experience a greater rate of divergence in the A than in the D homoeologs (Supplementary Dataset 2b). This relative acceleration of A-homoeolog divergence is mirrored in lineage-specific rate tests; the Gh/Gt clade including Upland cotton has the fastest evolving A homoeologs and the slowest evolving D homoeologs among five polyploids. These results demonstrate pervasive lineage-specific rate heterogeneities between subgenomes and among different polyploid cottons.

We examined patterns of gene loss and gain using 4,369 single-copy orthologs (SCOs), which are present in both diploids and in one or more allotetraploids (Extended Data Fig. 4c). Analysis of gene loss and gain among these basally shared homoeologs in the five polyploid lineages showed the highest level of net gene loss between the initial polyploidization and Gm, with threefold higher levels in the A subgenome (547 net gene losses) than in the D subgenome (149). Other polyploids have fewer gene losses with no subgenomic bias.

Among the homoeologs shared by all five polyploid species (Fig. 2a), the number of genes under positive selection (Ka/Ks values > 1) is the highest (3,200–3,300) in Gm with the longest branch relative to others, and the lowest between Gb and Gd (~1,100), the most recently diverged polyploid clade (Supplementary Dataset 3). Across different polyploid lineages, 10–20% more D homoeologs are under positive selection than A homoeologs, suggesting a concerted evolutionary impact on subgenomic functions in all polyploid species.

Fig. 2: Gene family expansion and contraction in cultivated and wild allotetraploid cotton species.
figure 2

a, Venn diagram representing the shared orthologous groups (orthogroups) between cotton species. No species-specific orthogroups were identified by Orthofinder (Methods). b, Venn diagram of R-gene family expansion and contraction in these five species. c,d, Flower morphology of two cultivated (c) and three wild (d) polyploid cotton species. Scale bars, 30 mm. e,f, Gene Ontology (GO)-term enrichment of the shared genes in polyploid domesticated cottons (e) and wild species (f).

Source data

Genomic diversity among five polyploids

The two subgenomes in each of the five polyploid species are highly conserved at the chromosomal, gene content and nucleotide levels (Fig. 1b and Extended Data Fig. 1). The D subgenomes have fewer and smaller inversions than the A subgenomes (Fig. 1c), as reported for Gh25, except for a few small inversions in D10 of Gt–Gm and Gm–Gb and D12 of Gd–Gt–Gm. This level of structural conservation is similar to some polyploids such as wheat7 and Arabidopsis suecica34, but is different from others such as B. napus10, peanut35 and T. miscellus11, which show rapid homoeologous shuffling.

The genomic conservation is extended to gene order, collinearity and synteny (Fig. 1c). Among the annotated genes (74,561–78,338), 56,870 orthologous groups or 65,300 genes (32,650 homoeologous pairs) (84–88%) are shared among all 5 species (Fig. 2a and Supplementary Dataset 1f).

The number of SNPs is in the range of 4–12 million (1.7–5.2 SNPs kb−1) or 0.19–0.53% among 5 polyploid genomes (Supplementary Dataset 4 and Supplementary Note). Gm has the highest SNP level (0.53%) relative to the other 4 species, with the lowest between the most recently diverged species Gb and Gd (~0.19%). Similar trends of indels range from ~5.55 Mb (~0.76%) in Gm–Gt to ~3.35 Mb (~0.34%) in Gb–Gd (Extended Data Fig. 1 and Supplementary Dataset 5). The level of overall variation of SNPs and indels among cotton species is low, comparable to natural variation (3.5–4.1 SNPs kb−1) between Brachypodium accessions28 but lower than that (~7.4 SNPs kb−1) for subspecies of rice36. SNPs are more frequent in pericentromeric regions, while indel distributions coincide with gene densities (Fig. 1b and Extended Data Fig. 1).

TE exchanges between two subgenomes that equilibrate the genome-size variation

The size difference between the Ga (~1.7 Gbp) and Gr (~0.8 Gbp)20,21 genomes is preserved in the respective A and D subgenomes of the 5 allotetraploid species (Fig. 3a). The A subgenome consists of a substantial amount of repetitive DNA in centromeric and pericentromeric regions (Fig. 3b). However, the A subgenome has 4.0–5.9% lower repetitive DNA content than the A-genome diploid (Ga), whereas the D subgenome has 1.5–2.9% higher content than the D-genome diploid (Gr) in Gh (Fig. 3c) and the other 4 species (Extended Data Fig. 5a). Consistently, the D subgenome has 10–20% more long terminal repeat (LTR) TEs than the D-genome diploid, while the A subgenome has 3–11% fewer LTRs than the A-genome diploid. These changes in subgenomic TEs may account for slight genome downsizing (Table 1) and genome-size equilibration following allopolyploidy in all five species, suggesting that the ‘evolutionary tape’ is replayed across polyploid lineages.

Fig. 3: Genomic diversification of A and D subgenomes in five allotetraploid cotton species.
figure 3

a, Chromosome painting of Gh using DNA probes to label telomeres (green) and 25S ribosomal DNA (red); DNA is stained by 4′,6-diamidino-2-phenylindole (DAPI, blue). The A (lower half) and D (upper half) homoeologous chromosomes are separated and rearranged into a tree shape. b, Pairwise comparison (dot plots) of 18-nucleotide sequences between the Gh and Gb homoeologous chromosomes D01 (top) and A01 (bottom) using Genome Pair Rapid Dotter (Gepard) plot analysis (Methods). The blue arrowheads indicate approximate centromeric locations. Genomic length positions are shown in each plot. c, Cumulative percentages (y axis) of 20-nucleotide sequences and their frequencies (x axis) in the A and D subgenomes of Gh relative to Ga (A) and Gr (D). d, The divergence time (Ma) of TEs (Copia and Gypsy) in the A and D subgenomes relative to their A- and D-genome-like diploids, Ga.A and Gr.D, respectively. The divergence time was estimated using the synonymous substitution rate (r) of 3.48 × 10−9 (Methods and Supplementary Note).

Source data

Copia- and Gypsy-like TEs are the most abundant LTRs in the Gh genome25. Estimates indicate that divergence of 5.6% (Gt) to 15.5% (Gh) and 39.7% (Gb) LTRs occurred during polyploid diversification (<0.6 Ma; Extended Data Fig. 5b–f). Since polyploid formation, LTRs increased substantially in the D subgenome of all five polyploids (Fig. 3d). The results indicate activation of LTRs in the D subgenome following polyploidization or movement of LTRs from the A to D subgenome37. Indeed, some Copia- and Gypsy-like elements are present in the D subgenome but absent in the extant D-genome diploid (Extended Data Fig. 5g).

Gene family diversification

The domesticated (Gh and Gb) and wild (Gm, Gt and Gd) cotton species share 417 (403) and 464 (359) unique genes (orthogroups) in respective groups (Fig. 2a), and no species-specific orthogroups are identified, although they possess distinct phenotypic traits such as fiber length (Fig. 1a) and flower morphology (Fig. 2c,d). The unique genes in the two domesticated cottons are over-represented in biological processes such as microtubule-based movement and lipid biosynthetic process and transport in the domesticated cottons (Fig. 2e; P < 0.05), reflecting the traits related to fiber development and cottonseed oil. Moreover, many of these genes are under positive selection and overlap regions of domestication traits including fiber yield and quality in Upland cotton38 (Supplementary Dataset 6). The unique genes in all three wild polyploid species, however, are enriched for pollination and reproduction (Fig. 2f), suggesting a role of these genes in reproductive adaptation in natural environments.

Plants have evolved an intricate innate immune system to protect them from pathogens and pests through intracellular disease-resistance (R) proteins as a defense response39. Among the R genes (Methods and Supplementary Note), each species has its unique R genes with very few genes shared between species (Fig. 2b and Supplementary Dataset 7), despite 5 wild and cultivated species sharing a core R-gene set (271), suggesting extensive diversification of R genes during selection and domestication. This is in contrast to a shared set of unique genes (related to fiber and seed traits) between the two cultivated species and the other shared set (related to reproductive and adaptive traits) among the three wild species (Fig. 2a)

Between the two subgenomes, the D subgenome has higher numbers of R genes (7.8%) than does the A subgenome (P = 0.0126, Student’s t-test; Supplementary Dataset 7). Using the published data40, we found expression induction of ~96% of 291 and 384 predicted R genes in the A and D subgenomes, respectively, by bacterial blight pathogens; 19 in D and 7 in A are upregulated at significant levels (error corrected, FDR = 0.05 and P < 0.001, exact test), while a similar trend of R-gene expression is observed after the reniform nematode attack (Supplementary Dataset 8), suggesting a contribution of the D-genome species to disease-resistance traits.

Gene expression diversity

In the five allotetraploid species sequenced, gene expression diversity is dynamic and pervasive across developmental stages and between subgenomes (Supplementary Dataset 9). Principal component analysis shows clear separation of expression between developmental stages (PC1) and between subgenomes (PC3; Extended Data Fig. 6a), with more D homoeologs expressed than A homoeologs in most tissues examined (Extended Data Fig. 7), consistent with higher levels of tri-methylation of Lys 4 on histone H3 (H3K4me3) in the former than in the latter41. Notably, expression correlates more closely with the subgenomic variation than with tissue types, except for fiber elongation and cellulose biosynthesis, where subgenomic expression patterns are more closely correlated between Upland and Pima cottons (Extended Data Fig. 6b). This may suggest that domestication drives parallel expression similarities of fiber-related genes in the two cultivated species.

These differentially expressed genes in fibers may contribute to fiber development, as they show enrichment of GO groups in hydrolase and GTPase-binding activities (Extended Data Fig. 8a,b). Hydrolases are essential for plant cell wall development42, and Ras and Ran GTPases are implicated in the transition from primary to secondary wall synthesis in fibers43. Moreover, translation and ribosome biosynthesis pathway genes are enriched during fiber elongation in Upland cotton and during cellulose biosynthesis in Pima cotton, consistent with faster fiber development in Upland cotton and longer fiber duration in Pima cotton44.

Expression networks and m6A RNA in fibers

Gene expression diversity is also reflected by coexpression modules in fibers among four species (Supplementary Dataset 10 and Supplementary Note). These module-related genes show higher semantic similarities between domesticated cottons (Gh–Gb) than with two wild species (Gt and Gm). The modules include supramolecular fiber organization genes in Upland cotton and brassinosteroid signaling genes in Pima cotton, which could affect fiber cell elongation45. The two wild species have different biological functions and transcription factors enriched in fiber-related gene modules (Supplementary Dataset 11), which may account for the fiber traits that are very different from those of the domesticated species (Fig. 1a).

Transcriptional and post-transcriptional regulation, including the activity of small RNAs and DNA methylation, mediates fiber cell development46. Modification of m6A messenger RNA can stabilize mRNA and promote translation with a role in developmental regulation of plants and animals47. In Upland cotton, m6A peaks are found largely in the 5ʹ and 3ʹ untranscribed regions (Extended Data Fig. 8c) of 1,205 genes in developing fibers (Supplementary Dataset 12), at levels 7-fold more than in leaves (Extended Data Fig. 8d) (P < 0.002, Student’s t-test), while the number of expressed genes is similar in both tissues. Notably, both m6A-modified mRNAs and transcriptome data in the fibers target the genes involved in translation, hydrolase activity and GTPase-binding activities (Extended Data Fig. 8a). These results indicate that mRNA stability and translational activities may determine fiber elongation and cellulose biosynthesis when cell cycles arrest in fiber cells.

Recombination and epigenetic landscapes

Polyploidy leads to low genetic recombination, as observed in B. napus48, which may comprise bottlenecks for breeding improvement. To determine the recombination landscapes in polyploid cottons, we genotyped 17,134 SNPs using the new Gh sequence and the CottonSNP63K array49 and identified a total of 1,739 low-recombination haplotype blocks (cold spots) in Upland cotton using whole-genome population-based linkage analysis50 (Methods and Supplementary Note). These blocks (average ~678.9 kb with 8.4 SNPs) span 1.18 Gbp (~52%) of the genome, including ~58% and ~41% in the A and D subgenomes, respectively (Fig. 4a), and are dispersed among all chromosomes with large ones predominately near pericentromeric regions. Recombination is generally suppressed throughout haplotype blocks, in contrast to that in subtelomeric regions (Extended Data Fig. 9a).

Fig. 4: Low-recombination haplotype blocks and their stability and selection during breeding and domestication.
figure 4

a, Distribution of presence (red) or absence (blue) of low-recombination haplotype blocks (red) in each pseudo-molecule of the A (A01–A13) and D (D01–D13) subgenomes in Gh. Map positions (Mb) are indicated in the y axis. b, A low-recombination haplotype block near the pericentromeric region (~72 Mb) of chromosome A08 (bottom). The color indicates the coefficient of linkage disequilibrium (D′) from low (blue) to high (red) with the upper confidence bound (D′ = 0.90) for the recombination cutoff. The recombination rates (y axis; using locally estimated scatterplot smoothing (LOESS) regression, Methods and Supplementary Note) in Gb × GhF2 (blue), Gm × GhBC1F1 (green), Gt × GhBC1F1 (yellow) and the consensus (red) are shown above with the positions (Mb, x axis). Two elevated recombination events are detected in Gb × GhF2 (blue arrows) and one in Gm × GhBC1F1 (green arrow). c, The average percentage (%) of CG (circle), CHG (triangle) and CHH (cross) methylation in the recombination hotspots (red) and cold spots (blue) between Gb and Gh. The CHH methylation is clustered in the left lower corner, which is visible in an enlarged image (Extended Data Fig. 10a).

Source data

Chromosome A08 has 62 haplotype blocks, including an exceptionally large one (~72 Mb) (Fig. 4b). Interestingly, interspecific hybridization between different tetraploids can increase recombination rates in these regions. For example, in the Gb × GhF2 population, recombination rates increased more than 4–6 cM Mb−1 in the left region (29–30 Mb) and in two other regions in the same Gb × GhF2 population. Recombination rates were also increased in the Gm × GhBC1F1 population (Fig. 4b). Similar increases were observed in the homoeologous D08 low-recombination haplotype blocks in the Gb × GhF2 population. Moreover, these haplotype blocks of either parent segregated with expected ratios within the population of Gh × GmBC2F1 (Extended Data Fig. 9b) or Gh × GtBC3F1 (Extended Data Fig. 9c). These data suggest the stability and selection of these haplotype regions during domestication and breeding.

Notably, genome-wide recombination cold spots (haplotype block) and hotspots (no haplotype block) correlated with the DNA methylation frequency at CG, CHG (H = A, T or C) and CHH sites in the cultivated allotetraploids Gh and Gb (Pearson r = 0.994; Fig. 4c and Extended Data Fig. 10a,b), with higher methylation frequencies in the cold spots than in the hotspots (analysis of variance (ANOVA), P < 1-10e). The data support the role of DNA methylation in altering recombination landscapes, as reported in Arabidopsis51,52. Consistent with this notion, DNA methylation changes that are induced in the interspecific hybrid (Ga × Gr) are also largely maintained in the five allotetraploid cotton species, creating hundreds and possibly thousands of epialleles, including the ones responsible for photoperiodic flowering and worldwide cultivation of cotton53.

Moreover, recombination events in all three interspecific crosses (Gb × GhF2, Gm × GhBC1F1 and Gt × GhBC1F1) correlated negatively with the average numbers of strongly connecting sites (intensity > 5) (P < 8.842 × 10−16) and their connection intensities (P < 7.26 × 10−12) of the Hi-C chromatin matrix (Pearson r = −0.874; Extended Data Fig. 10c). Recombination hotspots have fewer but more intense chromatin interactions within short distances, while the cold spots tend to have more but weaker interactions in long distances (Extended Data Fig. 10c,d). For example, 2 hotspots and 9 cold spots in the A08 region (Extended Data Fig. 10d), including 7 cold spots spanning ~32 Mb correlated with weak Hi-C intensities and DNA hypermethylation (Extended Data Fig. 10e). These data indicate that DNA hypermethylation and weak chromatin interactions interfere with recombination events in polyploid cottons.


Despite wide geographic distribution and diversification, five allotetraploid cotton genomes have largely retained the gene content and genomic synteny relative to respective extant diploids. This level of genome stability is in contrast to rapid genomic changes observed in some newly formed allotetraploids such as B. napus10 and T. miscellus11. However, in cultivated canola, the two subgenomes are relatively undisrupted8, probably because the extant parental species existing today to make new tetraploids10 may be different from the ones that formed cultivated canola ~7,500 years ago16 and likely became extinct. In addition, all five cotton polyploid species have a monophyletic origin, which is similar to the origin of wild and domesticated tetraploid peanuts54, but different from recurrent formation of Tragopogon tetraploids55. Notably, since polyploid formation 1–1.5 Ma, the evolution of 2 subgenomes in each of the 5 allotetraploid cotton species does not exhibit a simple asymmetrical pattern, as reported in Upland cotton25. Instead, the two subgenomes have diversified and experienced novel heterogeneous evolutionary trajectories, including partial equilibration of subgenome size mediated by differential TE exchanges, pervasive evolutionary rate shifts, and positive selection between homoeologs within and among lineages. These features present in all five allotetraploid species suggest that the ‘evolutionary tape’ is replayed during polyploid diversification and speciation.

Among the five allotetraploid genomes, no species-specific orthologs were identified, except for one set of the unique genes related to fiber and seed traits in the two domesticated cottons and another set of the unique genes for reproduction and adaptation in the three wild polyploid species. However, R-gene families have rapidly evolved in each allotetraploid and extensively diversified during selection and domestication. These genomic diversifications have been accompanied by dynamic and prevalent gene expression changes during growth and development between wild and cultivated polyploid species, including parallel gene expression, coexpression networks and m6A mRNA modifications in fibers of the cultivated species. Remarkably, polyploid cotton genomes show recombination suppression or haplotype blocks, which correlate with altered epigenetic landscapes and can be overcome by wild introgression and possibly epigenetic manipulation. This finding is contemporary to the discovery of the Ph1 locus that inhibits pairing of homoeologous chromosomes in polyploid wheat56,57. The recombination suppression may help maintain a repository of epigenes or epialleles that were generated by interspecific hybridization accompanied by polyploidization and could have shaped polyploid cotton evolution, selection and domestication53. These conceptual advances and genomic and epigenetic resources will help improve cotton fiber yield and quality as a sustainable alternative to petroleum-based synthetic fibers. Modifying epigenetic landscapes and using gene-editing tools may also overcome the limited genetic diversity within polyploid cottons. These principles may facilitate future efforts to concomitantly enhance the economic yield and sustainability of this global crop and possibly other polyploid crops.


Plant materials

G. hirsutum L. acc. TM-1 (1008001.06), G. barbadense L. acc. 3-79 (1400233.01), G. tomentosum L. (7179.01,02,03), G. darwinii L. (AD5-32, no. 1808015.09) and G. mustelinum L. (1408120.09, 1408120.10, 1408121.01, 1408121.02, 1408121.03) were grown in a greenhouse in College Station at Texas A&M University. Young leaves were collected for preparation of high-molecular-weight DNA using a published method58. Total RNA was extracted from leaf, root, stem, square, cotyledon, hypocotyl, meristem, petal, stamen, exocarp, ovule (0, 3, 7, 14, 21 and 35 days post anthesis (DPA)) and fiber (7, 14, 21 and 35 DPA) tissues in Gh; from leaf, root, stem, square, cotyledon, flower, ovule (14 DPA) and fiber (14 DPA) tissues in Gb; from leaf, root, stem, square, cotyledon and fiber (14 DPA) tissues in Gm; from leaf, root, stem, square, flower, ovule (0, 7, 14, 21 and 28 DPA) and fiber (7, 14, 21 and 28 DPA) tissues in Gt; and from leaf, root and stem tissues in Gd. Two or three biological replicates were used for RNA-seq and m6A RNA-seq analyses.

Genome sequencing and assembly

Sequencing reads were collected using Illumina HiSeq and NovaSeq and PacBio SEQUEL and RSII platforms. We sequenced and assembled five Gossypium genomes using high-coverage (>74×) single-molecule real-time long-read sequencing (Pac Biosciences). A total of six Illumina libraries were sequenced using the HiSeq platform, and two libraries were sequenced using NovaSeq. Initially, all five species were assembled using MECAT59 and subsequently polished using long reads, as well as Illumina reads. Gb and Gh were polished using QUIVER60, while Gd, Gt and Gm were polished using ARROW60. Ten Hi-C libraries were sequenced for five cotton genomes (two for each species). The total amount of Illumina sequenced for all 5 species (Supplementary Dataset 1) is 4,361,212,302 reads for a total of 286.4× of high-quality Illumina bases. A total of 105,182,984 PacBio reads were sequenced for all 5 genomes with a total coverage of 439.61×.

Chromosome integration of Gb and Gh leveraged a combination of published Gh synteny and Hi-C scaffolding. A total of 148,239 unique, non-repetitive, non-overlapping 1-kb sequences were extracted from the published Gh genome25 and aligned to the Gh and Gb MECAT assemblies. Misjoins in the MECAT assembly were identified, and the assembly was scaffolded with Hi-C data using the JUICER pipeline61. Small rearrangements to both genomes were made using the JUICEBOX interface62. Finally, a set of 5,275 clones (474.3 Mb total sequence) were used to patch remaining gaps in the Gh assembly. A total of 626 gaps were patched resulting in 1,871,050 base pairs (bp) being added to the assembly. Gd and Gm were integrated into chromosomes using Gb (3-79) synteny, whereas Gt was integrated using the Gh release assembly version 1!info?alias=Org_Ghirsutum_er. Final refinements to the Gt assembly were made using the JUICER/JUICEBOX pipeline61. In all five of the assemblies, care was taken to ensure that the telomere was properly oriented in the chromosomes, and the resulting sequence was screened for retained vector and/or contaminants. Genome annotation and gene prediction procedures are provided in the Supplementary Note.

Dot plots (pairwise comparisons) were generated using Gepard version 1.30 (ref. 63). The input data consist of 2 FASTA files, as well as the appropriate flags (-seq1 FASTA_FILE_1 -seq2 FASTA_FILE_2 -matrix edna.mat -zoom 65000 -word 18 -lower 0 -upper 20 -greyscale 0 -format png), with the -zoom flag from 65,000 (D subgenome) to 119,000 (A subgenome). The edna.mat file is part of the Gepard version 1.30 release. As a rule of thumb, this factor is generated by dividing the number of bases of the input FASTA file by 1,000. The output from the Gepard command is a PNG image file.

Procedures for the analysis of SNPs and indels are provided in the Supplementary Note.

Comparative analysis with published assemblies

Assessment of genome completeness

We evaluated the genome assembly completeness by k-mer masking (24-nucleotide) reciprocally between Gh (TM-1)22 and Gh (TM-1, this study) and between Gb (Hai7124)22 and Gb (3-79, this study). The unmasked contiguous sequences of the unshared sequence were extracted into a FASTA file and analyzed using FASTA statistics. BBMap ( and Custom Python scripts (Supplementary Note) were used for this analysis.

Genome comparisons using Hi-C data

The Hi-C libraries IKCF (Gh) and ILDE (Gb) were aligned to published Gh and Gb reference genomes using BWA-MEM64. Heatmaps were generated using the JUICER-pre command, and visualized using JUICEBOX62. Inversions and rearrangements were further identified using JUICEBOX.

Analysis of chromosomal collinearity, structural rearrangements and gene family composition between reference assemblies

Published Gh and Gb assemblies22 were aligned to the assemblies generated in this study using Minimap2 (ref. 65) with the parameter setting ‘-ax asm5 --eqx’. The resulting alignments were used to identify structural rearrangements and local variations using SyRI66. The gene copy numbers and gene families between assemblies were identified using OrthoFinder67 based on all annotated protein-coding sequences.

Analysis of evolutionary rate changes and gene gain and loss

Evolutionary rate changes in subgenomes of allopolyploid cotton during diversification

Rates of evolution for each subgenome of each species across the phylogeny were calculated using pairwise p-distances for the same 17,136 orthologs in all 5 polyploid species (Extended Data Fig. 4a). The distribution of p-distances between each species was compared for both subgenomes using a one-tailed Wilcoxon signed rank test and Bonferroni correction for multiple testing. Differences in evolutionary rates between the subgenomes within each species were evaluated using a modified relative rate test whereby p-distance distributions were compared for both subgenomes to determine which had the greater p-distance (that is, higher inferred rate). Differences in subgenome evolutionary rates among lineages were estimated using a modified relative rate test that again used the Wilcoxon signed rank test with the p-distances of 17,136 genes, here comparing p-distances between two species relative to an outgroup species. This test was repeated for all possible pairs of tip and outgroup combinations. We also summed the total number of differences contained within all orthologs between each pairwise set of species, excluding all sites in which any of the orthologs contained a gap sequence (Supplementary Dataset 2a). Chi-square tests were used to determine the significance of these total substitution counts (Supplementary Dataset 2b).

Analysis of gene loss and gain after polyploid cotton formation

A total of 32,622 groups of SCOs were identified between subgenomes of all 5 allopolyploids and the diploids Gr and Ga (Extended Data Fig. 4c). Of those, the 4,369 SCO groups that were present in both diploid species but absent in at least 1 allopolyploid subgenome were evaluated for gene losses specific to allopolyploids. The list of SCO groups was converted into a binary matrix of gene occurrence and mapped onto the inferred phylogeny of ten allopolyploid subgenomes (with five taxa each in the At- and Dt-subgenome clades, rooted by the respective diploid progenitors). Using a likelihood‐based mixture model assuming predominantly gene losses over gains and stochastic mapping implemented in GLOOME68, both the total number of gene gains and losses per branch and the associated probability of each event across the phylogeny were estimated.

Identification of homoeologs under selection

The homoeolog pairs of five species were used for estimating non-synonymous/synonymous (Ka/Ks) values. Every pair of the sequences were aligned using the MUSCLE alignment software69 and then transferred to the AXT format for identifying positively selected genes (Ka/Ks > 1) using the KaKs calculator70. Positively selected genes in A and D homoeologs were compared pairwise among 5 species (Supplementary Dataset 2).

Analyses of repetitive sequences and TEs

Pairwise comparison of 18-nucleotide sequences between homoeologous chromosomes was performed by Gepard plots63. Analysis of the k-mer content of all of the genomes was conducted by LTR-harvest71 according to the manual. The whole-genome sequences were suffixed first and then indexed using the seed length 20. The frequency of individual 20-nucleotide sequences was estimated using in-house Perl scripts. This analysis was applied to the two diploid cotton species, Ga and Gr, and the five tetraploid allopolyploids, with the A or D subgenome examined separately. The software LTR-harvest71 and LTR-finder72 was used for identifying full-length LTR retrotransposons. The identification parameters were as follows. For LTR-harvest: overlaps best -seed 20 -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. For LTR-finder: -D 15000 -d 1000 -L 7000 -l 100 -p 20 -C -M 0.9. The two datasets were integrated to remove false positives using the LTR-retriever packages73. The insertion time was estimated using the formula T = Ks/2r, where Ks is the divergence rate and r (3.48 × 10−9) is the substitution rate in cotton17.

Full-length TE sequences were extracted from each of the seven species and were used to build a TE database; the cd-hit software74 was applied to remove redundancies through self-sequence similarity tests, and sequences with identity > 90% were grouped into the same cluster. A cluster present in only one species was defined as a species-specific TE cluster, and those present in more than one species were considered shared TE clusters. A total of 98,794 full-length LTRs were identified in all 7 cotton species and grouped into 20,583 clusters for analysis of their origins in Ga, Gr, and the A and D subgenomes in 5 allotetraploids.

R-gene family and expression analysis in response to pathogen treatments

We detected nucleotide-binding site, leucine-rich repeat (NBS–LRR) motifs with the pfamscan tool75 that uses the hidden Markov model search tool (HMMER) version 3.2.1 (ref. 76) by searching primary protein-coding transcripts of each of the 5 allotetraploid cottons against the raw hidden Markov model for the NB-ARC-domain family downloaded from Pfam (PF00931). Identified NBS–LRR protein-coding genes for each of the allotetraploid cottons were further analyzed for amino-terminal (TIR/coiled-coil/other) and other functional domains by searching them against the Pfam-A hidden Markov model with the PfamScan tool and HMMER version 3.1 (ref. 76) with default settings (Supplementary Note). Short-read sequencing data for bacterial blight were downloaded from the Sequence Read Archive from the NCBI Bioproject accession PRJNA395458 (ref. 40). Reniform nematode sequence data were downloaded from the NCBI Bioproject accession PRJNA269348. Sequence data were aligned to the 653 predicted R genes from the Gh version 2.0 (this study) with Bowtie2 version and filtered for true-pair alignments. Fragments per kilobase million (FPKM) and read counts per million were determined with RSEM version 1.3.0. Differentially expressed R genes were determined with edgeR77 using false discovery rate (FDR)-corrected P values of 0.05. Of the 291 A-subgenome and 384 D-subgenome predicted R genes, we found FPKM expression profiles (>1) for at least 1 condition in 281 and 372 of the A- and D-subgenome predicted R genes, respectively. Similarly, in response to reniform nematode challenge in Gh, 274 of 291 A-subgenome and 370 of 384 D-subgenome predicted R genes were expressed at the FPKM level (>1) for at least 1 of the 4 conditions tested.

RNA-seq library construction, sequencing and data normalization

Total RNA was extracted from leaf, root, stem, square, flower, ovule and fiber samples from Gh, Gb, Gt, Gm and Gd species (2 replicates each for 124 samples; Supplementary Dataset 9), using PureLink Plant RNA Reagent (ThermoFisher). After DNase treatment, RNA-seq libraries were constructed using an NEBNext Ultra II RNA Library Kit (NEB), and 150-bp paired-end sequences were generated using an Illumina Hiseq 2500.

Paired-end sequence data were quality trimmed (Q ≥ 25) and reads shorter than 50 bp after trimming were discarded. Sequences were then aligned to respective allotetraploid cotton genomes and counts of reads uniquely mapping to annotated genes were obtained using STAR (version 2.5.3a). Outliers among the biological replicates were verified on the basis of the Pearson correlation coefficient, r2 ≥ 0.85. Fragments per kilobase of exon per million (FPKM) fragments mapped values were calculated for each gene by normalizing the read count data to both the length of the gene and the total number of mapped reads in the sample and considered as the metric for estimating gene expression levels78. Normalized count data were obtained using the relative logarithm expression (RLE) method in DESeq2 (version 1.14.1)79. Genes with low expression were filtered out, by requiring ≥2 RLE-normalized counts in at least 2 samples for each gene. Additional data for RNA-seq expression in fiber (28 DAP) tissue in both Gh and Gb were downloaded from the published data44 and processed as described above and in the Supplementary Note.

Statistical analysis of differentially expressed genes

To measure the gene expression differences between homoeologous genes in RNA-seq data, we used the DESeq2 package in R based on the negative binomial distribution (Supplementary Note). Only genes with log2[fold change] ≥ 1, Benjamini–Hochberg-adjusted P < 0.05 were retained. The comparison of highly expressed homoeologous gene pairs between subgenomes in different tissues was carried out using a binomial test (P < 0.05). GO enrichment was analyzed using topGO80, an R Bioconductor package with Fisher’s exact test; only GO terms with P < 0.05 (FDR < 0.05) were considered significant.

Principal component analysis and correlation coefficient analysis

To visualize subgenome and tissue expression relatedness, we used categorized gene expression values. These expression values were averaged across replicates and log2-transformed. Principal component analysis employed singular value decomposition via the prcomp function in R81. Categorized gene expression values were used in this analysis. Pearson’s correlation coefficients were determined and hierarchical clustering was carried out using the Euclidian distance and complete linkage method.

m6A RNA-seq data analysis

m6A RNA-seq libraries were constructed using a modified protocol as previously described82. Briefly, total RNA was extracted from young leaf and fiber tissues at 7 DPA (2 replicates each) from Gh by using PureLink Plant RNA Reagent (ThermoFisher). mRNA was collected from total RNA by the Oligotex mRNA mini kit (QIAGEN), fragmented and pulled down using an m6A antibody, followed by library construction using the NEBNext Ultra II RNA Library Kit (NEB) without polyA tail selection. Fragmented mRNA-seq libraries (control; input) and m6A RNA-seq libraries (IP) were sequenced using an Illumina Hiseq 2500 and 150-bp reads. Illumina reads were mapped to the Gh genome using Tophat 2.1.1 (ref. 83), and the uniquely mapped reads were used to identify m6A peaks with the Bioconductor package exomePeak84 (Supplementary Dataset 12).

GO terms were extracted from the GeneAnnotation_info.txt file. Identified m6A peak genes were analyzed by the Bioconductor package topGO80 to identify significantly over-represented GO terms (P < 0.0001). The location of RNA (5ʹ UTR, CDS or 3ʹ UTR) for each m6A RNA-seq read (both input and IP) was identified using the intersect function of Bedtools85. Single, double and triple asterisks indicate statistical significance levels of P < 0.05, P < 0.01 and P < 0.001, respectively (Student’s t-test).

We extracted the gene expression data for Gh leaf and fiber at 7 DPA corresponding to m6A peak genes. ‘All’ refers to the expression level of all identified homoeologous genes in the leaf and fiber samples, while ‘peak’ corresponds to the expression level of the identified m6A peaks for the genes in leaf (161 genes) and fiber (1,205 genes) samples. Single, double and triple asterisks indicate statistical significance levels of P < 0.05, P < 0.01 and P < 0.001, respectively (Student’s t-test).

Fluorescence in situ hybridization of A and D homoeologous chromosomes

Procedures for the preparation of metaphase chromosomes in Gh and fluorescence in situ hybridization were adopted from a published protocol86, with a modification that the cotton root tips were pretreated with cycloheximide (25 ppm) for 3 h at room temperature. The 25S rDNA fragment was obtained from Arabidopsis87 and originally provided by R. Hasterok from Poland. Synthetic oligonucleotides for forward and reverse plant telomeric sequences were PCR-amplified and products were labeled by nick translation to create probe to detect telomeres88.

Genotyping and recombination rate analyses

Genotyping data representing an improved cotton panel of 257 Gh accessions were acquired from a previously published diversity analysis49 utilizing the CottonSNP63K array89. The genotyping data in 2 segregating populations included 18 lines each representing 1 family of a Gh × GmBC2F1 population and 33 lines each representing 1 family of a Gh × GtBC3F1 population. SNPs with a minor allele frequency greater than 5% and that had less than 10% missing data were retained. Genotyping data were further filtered for homeo-SNPs that occur due to intragenomic sequence identity89. Array ID sequences were aligned to the Joint Genome Institute Gh version 2.0 sequence assembly using BLASTn90 (version 2.7.1+) with a minimum e-value cutoff of 1 × 10−10. Homoeologous alignments were corrected for using previously published SNP segregation data89,91, as well as interspecific, bi-parental linkage mapping populations from their respective Gh × GmBC1F1 and Gh × GtBC1F1 initial mapping populations. Genotyping data were then imputed and phased using Beagle (version 4.1)92, and genotypes were converted to ABH format to distinguish genotypic parentage.

It is notable that erroneous SNP calling is a common problem in polyploids and especially in the AD-genome allotetraploid cotton because of homoeologous and paralogous sequences. This issue has been addressed through several methods89,93,94. In this study, we used the published method89 to avoid erroneous genotype calling and to provide accurate chromosome-specific and homoeologous haplotype structure. Furthermore, we used a historical estimation of recombination95, as shown in the haplotype structure using confidence intervals, as well as in two segregating populations, which led to the accurate estimates of recombination rates between parental alleles using linkage disequilibrium analysis95. The haplotype block partitioning was conducted with PLINK50 (Supplementary Note).

The recombination map for chromosome A08 of Gh was developed using 4 SNP-based genetic maps, including 3 of interspecific crosses between Gb × Gh (F2, n = 195), Gt × Gh (BC1F1, n = 85) and Gm × Gh (BC1F1, n = 59) and 1 consensus map that was generated using 3 intraspecific populations91. All genetic maps were aligned to the Joint Genome Institute Gh version 2.0 sequence assembly using the previously stated methods. Recombination map visualization was estimated using the R package MareyMap96 using the nonlinear LOESS method97, and the number of surrounding markers used to fit a local polynomial was 7.5% of the total number of markers per chromosome. Final map plotting was conducted using the R package ggplot2 (ref. 98). Localized recombination rates for chromosomes A08 and D08 were estimated using a 1-Mb non-overlapping sliding window with a minimum of 4 SNPs per window as a linear regression threshold using MareyMap.

DNA methylation analysis

Methylome sequencing data were downloaded from a published report53. In brief, methylC-seq reads of all allopolyploid cottons were mapped to genome sequences of Gh and Gb, respectively, using Bismark with the parameters (--score_min L,0,-0.2 -X 1000 --no-mixed --no-discordant)99. Only the uniquely mapped reads were retained and used for further analysis. Reads mapped to the same site were collapsed into a single consensus molecule to reduce clonal bias. Cytosine counts were combined into 1,000-bp windows using methylKit 1.2.4 (ref. 100).

The DNA methylation (CG, CHG and CHH) levels (percentage of methylated cytosines) and average Hi-C seq statistics (number of connections, intensity or interaction matrix, and distance) in each recombination spot were compared using custom Python scripts. The Pearson correlation coefficient (r) was estimated using singular value decomposition via the prcomp function in R81. Single, double and triple asterisks indicate statistical significance levels of P < 0.001, P < 1 × 10−5 and P < 1 × 10−10, respectively, using one-way ANOVA.

Chromatin conformation capture (Hi-C) sequencing analysis

Hi-C seq libraries were constructed using a previously described protocol101,102, with modifications. Briefly, young leaves from Gh, Gb, Gt, Gm and Gd (2 replicates each) and fiber samples from Gh were fixed in 1% formaldehyde, and nuclei were extracted. Fixed chromatin was digested with DpnII, filled in using biotin-14-dATP and ligated. The biotin-labeled DNA was extracted and pulled down to construct HiC-seq libraries. Sequencing of Hi-C seq libraries was performed using an Illumina Hiseq 2500 and 150-bp reads. Reads were mapped to respective genomes and analyzed by HiC-Pro103. The Hi-C read coverage is 205× for Gh, 45× for Gb, 36× for Gm, 22× for Gd and 17× for Gt. The Hi-C data were largely used to correct orientations and misalignments in the assemblies of contigs and scaffolds. For Gh, Hi-C data were used to generate chromatin connection heatmaps with the HiCPlotter ( Single, double and triple asterisks indicate statistical significance levels of P < 0.001, P < 1 × 10−5 and P < 1 × 10−10, respectively, using one-way ANOVA.

Reporting Summary

Further information on research design is available in the Nature Genetics Research Reporting Summary linked to this article.