Structural Variations Contribute to Subspeciation and Yield Heterosis in Rice

ABSTRACT

Yield heterosis has been extensively exploited in hybrid breeding, with intersubspecific hybrids often exhibiting the most pronounced effects. However, developing elite hybrids remains a laborious and time‐consuming process. The genetic basis of heterosis has been debated for over a century, hindered largely by the lack of high‐quality genomes. Here, we assembled genomes for 12 representative indica, intermediate type and japonica rice accessions. Using sequence variants of the Phr1 gene, we functionally validated two deletions responsible for phenol reaction variation between the subspecies. Comparative genomic analyses revealed extensive sequence variation among these inbred lines and highlighted the pivotal role of structural variants (SVs) in rice subspeciation. Importantly, the number of SVs between parental inbred lines significantly correlated with heterosis across 17 agronomic traits, with distinct correlation patterns for intra‐ and intersubspecific F₁ hybrids. We identified SVs associated with S5‐ORF5 and OsBZR1 and validated their function to heterosis for seed setting rate and yield heterosis, respectively, underscoring the importance of SVs in breeding intersubspecific hybrids. The genomic SVs altered gene expression and these transcriptional changes effectively explained the variance in heterosis. Furthermore, translocations outperformed other SVs and their heterozygous haplotypes exhibited heterosis over homozygous ones. Our findings establish SVs as pivotal drivers in subspeciation and highlight the overdominance model for harnessing rice heterosis.

1. Introduction

Heterosis refers to the phenomenon where heterozygous F₁ hybrids exhibit superior performance compared to their parental lines (Shull 1948), significantly enhancing crop yield and contributing to global food security (Hochholdinger and Yu 2025). Though several well‐known genetic models—including dominance (complementation of deleterious loci; Bruce 1910; Gu et al. 2023; Jones 1917; Wang et al. 2023), overdominance (heterozygosity of loci; East 1936; Krieger et al. 2010; Semel et al. 2006; Shull 1908, 1911), and epistasis (interaction of loci; Jiang et al. 2017; Powers 1944)—have been proposed to explain heterosis, the genetic mechanisms underlying crop heterosis remain debated. Over a century ago, Shull proposed the overdominance model, emphasising heterozygosity of ‘elements’, now always interpreted as genes or genomic sequences. He also suggested that the degree of heterosis correlates with the number of parental differential ‘characters’, now may be understood as genomic variants (Shull 1911). Consistent with the opinion of Shull, East postulated that heterosis arises from the action of nondefective alleles and is linked to parental genetic disparity (East 1909, 1936).

The advent of DNA marker technology in the late 20th century enabled investigations into the relationship between parental genetic disparity and heterosis in F₁ hybrids using various molecular markers (Stuber et al. 1992; Xiao et al. 1996, 1995; Zhang et al. 1996). Studies reported significant correlations between heterosis (or hybrid performance) and genetic distance or marker heterozygosity (Dan et al. 2014; Liu et al. 2014; Schrag et al. 2006, 2009, 2010; Xiao et al. 1996; Zhang et al. 1996). However, due to technology limitations, these studies relied on a small number of marker sets, unable to reflect genome‐wide sequence variation.

Advances in genome sequencing over the past two decades have dramatically improved the detection ability and completeness of genomic variants. High‐density single nucleotide polymorphism (SNP) and small insertion and deletion (InDel) markers (< 50 bp) have been widely used with various mathematical models to predict heterosis or hybrid performance in maize, rice, and wheat (Xu et al. 2014; Zhao et al. 2015, 2014). Sequencing thousands of rice hybrids revealed that the number of superior alleles correlated strongly with grain number per panicle (Huang et al. 2015). In‐depth sequencing analysis of 17 representative hybrid populations demonstrated that heterozygosity at heterosis‐associated loci underlay rice heterosis (Huang et al. 2016). The sequencing of 1143 indica inbred lines further underscored the importance of genomic variants in breeding superior hybrid rice (Lv et al. 2020). Based on sequencing results of 2839 hybrids and their 9839 segregation individuals, the authors found that the breeding of elite indica‐indica hybrids was a process of pyramiding favourable alleles from broadened genetic resources (Gu et al. 2023). They revealed widespread genomic complementarity in indica‐japonica heterosis (Gu et al. 2023). Moreover, genomic sequencing analysis of two test‐cross populations elucidated that rice heterosis was due to homozygous disadvantage at insufficient genetic background, rather than heterozygous advantage (Xie et al. 2022).

The contribution of large‐size genomic variants (≥ 50 bp), e.g., genome assembly‐dependent structural variants (SVs), to rice heterosis remains poorly understood. Recent progress in sequencing technologies, including Pacific Biosciences (PacBio), Oxford Nanopore Technologies and Hi‐C platforms (Li and Durbin 2024), has enabled the assembly of high‐quality genomes. Over the past 5 years, the completeness and accuracy of genome assemblies have been significantly improved, leading to the release of multiple high‐quality rice genomes (Qin et al. 2021; Shang et al. 2023, 2022; Sun et al. 2023). SVs, some of which originate from ancient accessions (Zhang, Wang, et al. 2022), are now recognised as key determinants of variations in agronomic traits. They have profoundly influenced rice subspeciation (e.g., indica vs. japonica), speciation (e.g., Oryza sativa vs. Oryza glaberrima ) and environmental adaptation (e.g., flooding, cold, osmotic stress; Li et al. 2021; Qin et al. 2021; Shang et al. 2022; Wei et al. 2024). Given the important roles of SVs in tomato improvement and maize heterosis (Alonge et al. 2020; Wang et al. 2023), SVs are likely also critical for rice heterosis. A comparative genomic analysis of 10 elite inbred lines, which are parents of widely planted, high‐yield hybrid rice, reveals that presence and absence variants (PAVs) constitute the predominant form of SVs (93.2%) and play important roles in gene expression heterosis (Zhang, Fu, et al. 2022). Genome assemblies of Tianfeng and Huazhan, the parents of an elite hybrid rice Tianyouhuazhan, enable the identification of PAVs in genes controlling vital processes such as plant growth (e.g., ARD1), disease resistance (e.g., Pit) and evolution (e.g., OsSh1) (Zhang, Wang, et al. 2022). Furthermore, high‐quality assemblies of parents of the super‐hybrid Y900 (Y58S and R900) uncover a copy number variant (CNV) encompassing the transcription factor OsWRKY89. This CNV alters the gene's expression, with Y58S showing higher levels than R900, while the F₁ hybrid Y900 exhibits an additive expression pattern (Sun et al. 2023). These findings collectively underscore the significance of SVs in driving heterosis in rice.

In this study, we selected 12 representative indica, intermediate type and japonica inbred lines and assembled their genomes for comparative analysis. We identified genomic variants, particularly SVs, and assessed their functional implications in rice subspeciation. Using these inbred lines as parents, we constructed a diallel‐cross population and analysed the relationship between parental genomic variants and heterosis of 17 agronomic traits. We identified SVs in S5‐ORF5 and OsBZR1 and validated their contributions to heterosis for seed setting rate and yield heterosis, respectively. In combination with gene expression data, we found that SVs drove transcriptional changes and that these effectively explained variance in rice heterosis. Furthermore, translocations outperformed other SVs and their heterozygous haplotypes conferred heterosis over the homozygous ones, supporting the overdominance model of heterosis in rice.

2. Results

2.1. Genome Assembly and Annotation of 12 Representative Rice Inbred Lines

To investigate the genetic mechanisms underlying rice heterosis, we selected 12 representative indica, intermediate type and japonica accessions from 223 inbred lines (ILs) based on previous genotyping data (Dan et al. 2024). These 12 ILs were equally clustered into two groups (Figures 1A and S1), with different types of genomic variants analysed from the reported Illumina sequencing data (Dan et al. 2024). To enable comparative analyses, we classified these ILs into indica and japonica groups. Among the selected ILs, seven—Yuetai B (YB), Mianhui725, Chenghui9348 (9311K), 610234, C418, Wuyunjing8 and Balilla—are well‐established cultivars known for high grain yield or their widespread exploitation in breeding elite inbred or hybrid rice (https://www.ricedata.cn; Figure S2). Other ILs show highly recognisable phenotypes, for example, large grain/primary branch/secondary branch number for Qianlijing, long panicle for R4115, and tall plant height for JR2 (Dan et al. 2024). These ILs represent phenotypically and geographically diverse varieties originating from China, India and Italy (Figure 1B).

FIGURE 1.

Genome assemblies and pan‐genome of 12 rice inbred lines. (A) Neighbour‐joining trees of 12 representative rice inbred line (ILs) based on genomic SNPs (left side), InDels (middle), and CNVs (right side) analysed from previously reported Illumina sequencing results. (B) Geographic origins and seeds of the 12 ILs. (C) Better‐parent heterosis for seed setting rate (BPH_SSR) of intra‐ and inter‐subspecific F₁ hybrids. The pooled average values of indica‐indica (i_i) and japonica‐japonica (j_j) F₁ hybrids were also included for comparison. Analysis of variance (ANOVA) with LSD in post hoc test (p value < 0.05) was performed. The number of F₁ hybrids are indicated. (D) Gene structures and neighbour‐joining tree of the 12 ILs based on gene sequences of Phr1. Nipponbare (japonica) and 9311 (indica) served as reference accessions. Red regions denote genomic variants, and dashed lines represent deletions. (E) Phenol reaction phenotypes of 14 ILs. The bar represents 1 cm. (F) Variation of gene family sizes in the pan‐ and core‐genome. (G) Core and noncore‐gene families across the 12 ILs. (H) Gene annotation ratios cross three databases. Three databases, including Pfam, GO and KEGG, were selected for annotation analysis.

Using a complete diallel‐cross design, we generated F₁ hybrids from the 12 ILs and calculated better‐parent heterosis (BPH) of 17 agronomic traits for the finally obtained 120 F₁ hybrids. We observed that F₁ hybrids derived from intersubspecies exhibited significantly reduced seed setting rate and increased plant height compared to those of intrasubspecific hybrids (Figures 1C and S3), which is consistent with known breeding challenges. Furthermore, heterosis of the intersubspecific hybrids showed significant differences from the pooled averages of indica–indica and japonica–japonica hybrids.

To generate genome assemblies, we obtained genomic data using single‐molecule real‐time (SMRT) technology on the PacBio Revio platform for the 12 ILs, achieving an average sequencing depth of 30.46× (Tables S1 and S2). We performed de novo assembly for two representative ILs, YB (indica) and Balilla (japonica), with the assistance of Hi‐C data, yielding an average sequencing depth of 127.80× (Figure S4 and Table S3). The genomes at the chromosome level of the remaining 10 ILs were assembled using YB and Balilla as reference genomes for indica and japonica subgroups, respectively. The average contig N50, scaffold N50 and total assembled genome sizes for final genome assemblies of the 12 ILs were 28.74, 31.75, and 396.36 Mb, respectively, with an average anchor ratio of 98.18%. The high quality and completeness of these genome assemblies were further confirmed by metrics such as average quality value of bases (63.06), re‐mapping rate of HiFi reads (99.51%), Benchmarking Universal Single‐Copy Orthologs (BUSCO) (98.61%) and long terminal repeats (LTRs) assembly index (22.13; Figure S5).

We annotated repetitive sequences (tandem repeats and transposable elements), protein‐coding genes and noncoding RNAs for YB and Balilla (Table S4). Repetitive sequences constituted 57.35% (YB) and 55.44% (Balilla) of their genomes, with transposable elements (TEs) comprising 55.09% and 53.10%, respectively. Among these TEs, DNA transposons and LTRs were the predominant types, accounting for 24.26%–25.67% and 24.9%–28.64% of the genomes, respectively. Three strategies, including homology‐based (using five previously reported genomes), de novo and transcriptome‐based (full‐length RNA sequencing) predictions, were integrated for the annotation of protein‐coding genes. This resulted in YB containing 44 332 genes and Balilla containing 43 706 genes. We revealed that 94.4% of the annotated genes in YB and 95.35% in Balilla have assigned functions in one of the following six databases: InterPro, Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), Swiss‐Prot, TrEMBL and nonredundant. Compared with the five established genomes (Nipponbare, 9311, Minghui63, Zhenshan97 and R498), these newly assembled genomes showed significant differences in the lengths of genes and coding sequences, highlighting their value to current rice genomic resources (Figure S6). Moreover, miRNA and rRNA were the two predominant types of non‐coding RNAs in the genome assemblies of both YB and Balilla, which accounted for 0.26%–0.28% and 0.34%–0.96% of the genomes, respectively.

To further validate assembly accuracy, we analysed Phr1, a gene underlying the phenol reaction phenotype between indica and japonica (Yu et al. 2008). The phylogenetic result of Phr1 sequences from the 12 ILs, along with 9311 and Nipponbare, displayed slightly altered clustering patterns in subspecies, with Qianlijing grouped with japonica (Figure 1D). Using the sequence of 9311 as a reference, Phr1 contained 24 SNPs and six InDels. Among these InDels, there existed three insertions (1–2, 6 and 7 or 14 bp) and three deletions (2, 18 and 29 bp), with only the 18 and 29 bp deletions located in exons. The result of the phenol reaction experiment revealed that seed hulls of five ILs without the 18/29 bp deletions immediately turned black after 1 day's treatment, while the other nine remained yellow or brown even after 124 days (Figure 1E). Our results showed the high assembly accuracy of these new genomes and confirmed important roles of the 18/29 bp deletions in the outcomes of the phenol reaction.

2.2. Pan‐Genome Assembly and Pan‐Structural Variants of 12 Inbred Lines

We constructed a pan‐genome for the 12 ILs, comprising 47 490 gene families (Figure 1F). Among the whole gene families, 26 556 (with 345 960 genes) were identified in all 12 ILs and defined as core‐gene families (55.92% of gene families and 70.70% of genes), 2037 (with 24 040 genes) identified in 11 ILs were defined as softcore gene families (4.29% of gene families and 4.91% of genes), 18 850 (with 118 436 genes) identified in 2–10 ILs were defined as dispensable families (39.69% of gene families and 24.20% of genes), and 47 (with 892 genes) identified in only one IL were defined as private gene families (0.10% of gene families and 0.18% of genes). The number of gene families increased with the addition of genomes, reaching a plateau at n = 10, indicating the representativeness of the selected ILs (Figure 1F). Unlike the trend observed in the pan‐genome, the number of core‐gene families decreased as the number increased, with the lowest point reached at n = 12, underscoring the specificity of each IL (Figure 1F). For the number of noncore‐gene families among these ILs, YB and Qianlijing had the highest and lowest amount, respectively, showing functional variations between ILs (Figure 1G). Furthermore, we analysed the proportion of genes with annotated functions in protein domain annotation (Pfam), GO and KEGG. The annotation ratios decreased progressively from core to softcore to dispensable genes (Pfam: 63.23%, 43.05% and 35.77%; GO: 69.55%, 56.17% and 51.02%; and KEGG: 27.97%, 13.12% and 10.26%), demonstrating the conservation of core genes amidst diversification (Figure 1H).

We identified genomic variants—especially SVs (presence/absence variants, copy number variants, translocations and inversions)—through comparing the reference genome YB to the other 11 ILs. Consistent with the patterns observed for SNPs and InDels (Figure 2A), both the number and size (about 57–168 Mb) of SVs gradually increased from indica to japonica (Figure 2B,C). All japonica ILs harboured more SVs than the indica. Notably, we detected two large inversions on chromosome 6 (inversion_6: 5.83 Mb) and 8 (inversion_8: 5.90 Mb; Figure 2D), which were between the six indica and six japonica or exclusive to JR2 (Figures S7–S11), respectively, indicating that inversions contribute to the evolutionary divergence of indica and japonica.

FIGURE 2.

Genomic variants of 12 inbred lines. (A–C) Number and size distribution of genomic variants in 11 inbred lines with genome assembly of YB as the reference. (D) Comparison of genomic profiles between YB and Balilla. (E) Variation of SVs in the pan‐ and core SVs with varied number of inbred lines. (F) Percentages of core, softcore, private and dispensable SVs. (G) Number of core, softcore, private, and dispensable SVs in 11 inbred lines. (H, I) Number and size of SVs. (J) Percentages of different types of repeats in SVs. The number and size are analysed.

Unlike the changing trend of pan‐genome, the number of pan‐SVs did not plateau even at n = 11, underscoring the necessity of including more accessions for a comprehensive exploration of SVs (Figure 2E). We assigned core, softcore, dispensable and private SVs using the same criteria as the pan‐genome. Among the total SVs, 878 (0.73%) identified in 11 ILs were classified as core SVs, while 1640 (1.36%) detected in 10 ILs were categorised as softcore SVs (Figure 2F). Additionally, 66 655 (55.28%) SVs identified in 2–9 ILs were classified as dispensable SVs, and 51 413 (42.64%) identified in only one IL were considered private SVs. Different from the trends of core, softcore and private SVs, the number of dispensable SVs increased from indica to japonica (Figure 2G), suggesting the roles that these dispensable SVs played in rice subspeciation. The length of different SV types varied (Figure 2H), with three large absence variants (AVs) detected on chromosome 6 (AV_6, 0.73 Mb), 8 (AV_8, 1.16 Mb) and 11 (AV_11 and 0.94 Mb; Figure 2I). Given that repetitive sequences accounted for more than 50% of the genomes, we analysed the distributions of SVs and found that these SVs were predominantly detected in DNA transposons and LTRs (Figure 2J), suggesting the regulatory functions of SVs in noncoding regions of the genomes.

2.3. Associations Between Structural Variants and Heterosis of 17 Agronomic Traits

We performed paired genomic comparisons and identified both unique and common genomic variants for each pair of ILs, representing different and shared variants, respectively (Figure 3A and Table S5). On average, unique genomic variants between any two ILs comprised 2 018 539 SNPs, 391 298 InDels and 40 082 SVs (17 690 AVs, 11 975 PVs, 6327 CNVs, 3941 translocations and 149 inversions). The number of these unique variants was highly correlated (Figure 3B), with the exception of translocation, inferring co‐locations of these variants (Figure S12). Approximately half of these variants located in intergenic regions (57.45% of SNPs, 50.42% of InDels and 50.56% of SVs), emphasising the significance of these regions in differentiating the ILs. In contrast to other SV types, the majority (77.94%) of unique inversions occurred in exonic regions. Furthermore, genetic relatedness of the 12 ILs assessed using unique SVs, particularly translocations, showed variation compared to those based on unique SNPs or InDels, indicating the distinct functional contributions of different variant types to genomic divergence (Figures 3C and S13).

FIGURE 3.

Correlations between parental genomic variants and heterosis of 17 agronomic traits. (A) Genomic comparisons of YB against the other 11 inbred lines (ILs). (B) Heatmap for the correlations between different types of unique genomic variants. (C) Genetic relatedness of 12 ILs based on unique SVs. (D) Correlation patterns between the number of parental genomic variants and heterosis of 17 traits. Parental unique and common variants were classified into different types according to their location or function. Spearman correlation analysis was performed with better‐parent heterosis (BPH) of the following traits: Grain number per panicle (GNP), tiller number per plant (TPP), seed setting rate (SSR), thousand‐grain weight (TGW), seedling stage tiller number (SSTN), elongation stage tiller number (ESTN), flowering stage tiller number (FSTN), maturation stage tiller number (MSTN), yield per plant (YPP), seedling stage plant height (SSPH), elongation stage plant height (ESPH), flowering stage plant height (FSPH), maturation stage plant height (MSPH), panicle length (PL), primary branch number (PBN), secondary branch number (SBN), and heading date (HD). The number besides each row indicates the number of significant correlations (p value < 0.05). (E) Comparison of correlation frequencies between unique and common genomic variants. p value indicates independent t‐test (weighted). (F) Correlations between the number of unique SVs and heterosis for seed setting rate in intra‐ and inter‐subspecific F₁ hybrids. A pair of reciprocal hybrids (F_{1: Mianhui725×Balilla} and F_{1: Balilla×Mianhui725}) with extremely low heterosis are shown. Mature seeds with black and white hulls are full and empty seeds, respectively. p values indicate Spearman correlation. (G) The number of unique SVs in intra‐ and inter‐subspecific F₁ hybrids. The two reciprocal hybrids are marked in red. ANOVA with LSD in post hoc test (p value < 0.05) was performed.

Correlation analysis between the number of parental genomic variants and heterosis of the 17 traits revealed divergent patterns for unique versus common variants (Figure 3D). According to the number of significant correlations, the unique variants exhibited stronger correlations with heterosis than common variants (Figure 3E). Meanwhile, no notable change in correlation patterns was observed between stop‐gain and stop‐loss variants in SNPs and InDels, synonymous and nonsynonymous SNPs or frameshift and nonframeshift InDels. Our findings emphasised the importance of genomic differences between the two parents in determining heterosis.

Though unique SVs accounted for only 1.99% and 10.24% of the total SNPs and InDels, respectively, their correlations with heterosis were comparable, probably due to their larger size. Among the five types of SVs, inversions, with a fewer number than translocations, showed stronger correlations with heterosis for plant height, inferring a size‐dependent influence of genomic variants on the heterosis of vegetative traits. Different from correlation patterns of heterosis for other yield‐related traits, the number of unique variants showed negative correlations with heterosis for the seed setting rate, showing reproductive isolation due to genomic disparity. Notably, we observed that an increase in unique SVs corresponded with a slight rise in heterosis for the seed setting rate of intrasubspecific F₁ hybrids, but a dramatic decrease in indica‐japonica F₁ hybrids (Figure 3F). The intersubspecific hybrids possessed significantly more unique SVs in their parents than their intrasubspecific counterparts (Figures 3G and S14), suggesting the roles of these SVs in the heterosis of intersubspecies. The unique SVs between the parental lines also exhibited stronger correlations with heterosis for tiller number in intersubspecific F₁ hybrids, showing the potential for enhancing grain yield through utilising heterosis for tiller number (Figure S15). However, the number of unique SVs displayed weak or no correlation with yield heterosis in intersubspecific F₁ hybrids, likely due to reproductive isolation‐driven reductions in seed setting rate.

2.4. Absence Variant of S5‐ORF5 Underlies Heterosis for Seed Setting Rate

Given the established role of the S5 loci in embryo‐sac fertility and seed setting rate of intersubspecific F₁ hybrids (Chen et al. 2008; Yang et al. 2012), we performed phylogenetic analyses of its three genes (S5‐ORF3, S5‐ORF4 and S5‐ORF5) in the 12 ILs (Figure 4A). Unexpectedly, two ILs—YB and Qianlijing—clustered together and were identified as wide compatibility varieties (WCVs). With the gene sequence of Balilla as the reference, we detected eight SNPs, three deletions (1, 10, and 42 bp), and one 94 bp AV in S5‐ORF5 (Figure 4A), which acts as the ‘killer’ gene. The 94 bp AV, which is linked with the 10 and 42 bp deletions, was exclusive to the two WCVs, potentially rescuing embryo‐sac viability in their intersubspecific F₁ hybrids.

FIGURE 4.

Contribution of AVs in S5‐ORF5 to heterosis for seed setting rate. (A) Neighbour‐joining tree of 12 inbred lines (ILs) based on S5 loci (three genes) and gene structure of S5‐ORF5. Genotypes of S5‐ORF3, S5‐ORF4, S5‐ORF5 for 12 ILs were assigned with previously reported rules. The ‘n’ represents wide compatibility varieties. The sequence of Balilla serves as the reference and sequence variants are indicated in red. (B, C) Comparisons of seed setting rate and better‐parent heterosis for seed setting rate in intra‐ and inter‐subspecific F₁ hybrids. Two wide compatibility varieties, namely YB and Qianlijing, were excluded for analysis. ANOVA with LSD in post hoc test (p value < 0.05) was performed. (D, E) Comparisons of seed setting rate and better‐parent heterosis for seed setting rate between intra‐ and inter‐subspecific F₁ hybrids with YB as a parent. (F, G) Comparisons of seed setting rate and heterosis for seed setting rate between intra‐ and inter‐subspecific F₁ hybrids with Qianlijing as a parent. p values in (D–G) indicate independent sample t‐test. (H) Panicles of inter‐subspecific F₁ hybrids with YB or Qianlijing as a parent. The letter f and m represent female and male parent, respectively. Seed setting rate of these F₁ hybrids are displayed in (D) and (F).

To investigate the contribution of the 94 bp AV to heterosis, we first compared the seed setting rate between intra‐ and intersubspecific F₁ hybrids lacking the WCV parentage. The seed setting rate of indica‐indica and japonica–japonica F₁ hybrids was similar, both significantly higher than those of indica–japonica F₁ hybrids (Figure 4B). The heterosis for seed setting rate was also lower in intersubspecific F₁ hybrids, compared to those of intrasubspecific F₁ hybrids and their pooled average values (Figure 4C). However, when YB or Qianlijing was used as a parent, the indica‐japonica F₁ hybrids achieved a seed setting rate as high as the indica–indica F₁ hybrids (Figure 4D–H). Importantly, heterosis for seed setting rate was also comparable between the inter‐ and intrasubspecific F₁ hybrids. We therefore identified YB and Qianlijing as two indica‐type WCVs and confirmed the important role of the 94 bp AV in heterosis for seed setting rate.

2.5. Validation of OsMADS1 in Yield Heterosis and Presence and Absence Variants of OsBZR1 in Subspecific Heterosis

The transcription factor OsMADS1, a MADS‐box protein, has been genetically demonstrated to regulate yield heterosis by altering mRNA metabolism in rice (Wang, Wang, et al. 2024). Mutants of OsMADS1 develop longer grains and exhibit a higher grain yield per plant than the wild type (Wang, Wang, et al. 2024). In F₁ hybrids generated from wild‐type and mutant lines, grain length and yield display intermediate values, indicating an additive effect in heterosis. Notably, heterozygous genotypes of OsMADS1 derived from divergent genetic backgrounds produce strong yield heterosis. To evaluate the role of OsMADS1 in our study, we conducted a phylogenetic analysis that identified three distinct haplotypes (haplotype_a, ‐b, and ‐c; Figure S16). With YB as a reference, we detected 39 SNPs, 9 insertions, 11 deletions, and one 82 bp AV within the gene (Figure S16). We subsequently compared heterosis for yield and seven yield‐related traits between F₁ hybrids with homozygous and heterozygous haplotype combinations. Consistent with prior findings (Wang, Wang, et al. 2024), F₁ hybrids with heterozygous haplotypes (haplotype_a/haplotype_c and haplotype_b/haplotype_c) of OsMADS1 showed significant yield increases compared to those with homozygous haplotypes (Figure S16), thereby validating the contribution of OsMADS1 to yield heterosis.

OsBZR1, a positive regulator of the brassinosteroid signalling pathway (He et al. 2002), regulates phenotypic differentiation between indica and japonica subspecies (Yang et al. 2025), including yield per plant, tiller number, grain number, thousand‐grain weight, panicle length and plant height. Clustering of the 12 ILs based on phenotypic data of these traits supported OsBZR1's role in subspecific differentiation (Figure S17). Given that sequence variants in the OsBZR1 promoter region—particularly a 4808 bp retrotransposon—are critical for its function, we extracted and compared OsBZR1 sequences in the 12 ILs. Phylogenetic analysis revealed that 11 of the 12 ILs clustered into distinct indica and japonica groups, with JR2 as an exception (Figure 5A). We thus classified the ILs into three haplotypes: haplotype_i (indica), haplotype_j (japonica) and haplotype_JR2 (divergent).

FIGURE 5.

Contributions of SVs in OsBZR1 to yield heterosis. (A) Neighbour‐joining tree of 12 inbred lines (ILs) based on sequences of OsBZR1. The 12 ILs were divided into three haplotypes. (B) Structure variations of OsBZR1 across 12 ILs. The sequence of Balilla serves as the reference. Single nucleotide polymorphism (SNP), insertion and deletion (InDel), absence variant (AV) and presence variant (PV) were indicated. (C) Correlation between expression level of OsBZR1 and better‐parent heterosis (BPH) for seedling length. Six pairs of reciprocal hybrids, which were classified as with high‐, low‐ and no‐BPH for seedling length, were adopted for correlation analysis. p value indicates Spearman correlation. (D–J) Comparisons of BPH for seven OsBZR1‐associated traits between heterozygous and homozygous haplotypes. The seven traits are yield per plant (YPP), tiller number per plant (TPP), grain number per panicle (GNP), thousand‐grain weight (TGW), panicle length (PL), maturation stage plant height (MSPH) and seed setting rate (SSR). Hap, i, and j represent haplotype, indica, and japonica, respectively. p values in (D–J) are for independent sample t‐test.

Using the sequence of Balilla as reference, we identified 82 SNPs, four insertions (1, 4, 9 and 12 bp), four deletions (1, 2, 2 and 8 bp), one 266 bp PV and six AVs (356, 523, 1008, 1486, 1743 and 4808 bp; Figure 5B). The 266 bp PV was exclusive to the six japonica accessions, while the 4808 bp AV was present in the six indica lines, indicating contributions of these SVs to subspeciation. The remaining five AVs, four of which were mainly derived from the 4808 bp AV, were unique to JR2.

To investigate the potential role of the retrotransposon in regulating gene expression, we analysed OsBZR1 expression levels across multiple ILs. Initially, a comparison between three indica‐group ILs (YB, 610234 and Qianlijing) and two japonica‐group ILs (C418 and Balilla) revealed higher OsBZR1 expression in the japonica‐group (Figure S18). We expanded this analysis to include six ILs from each subspecific group and confirmed that the expression levels of OsBZR1 were significantly elevated in japonica‐group lines (Figure S18), a finding consistent with previously reported patterns (Yang et al. 2025). We then analysed the relationship between expression levels of OsBZR1 and heterosis for seedling length in six pairs of reciprocal F₁ hybrids, which were divided into high‐, low‐, and no‐heterosis groups (Dan et al. 2024). We observed a significantly negative correlation, suggesting contributions of low expression levels of OsBZR1 to heterosis for seedling length (Figure 5C). When expression levels were compared between F₁ hybrids and corresponding parents, we detected overdominant effect, dominant effect and unclassifiable effects, indicating diverse inheritance patterns of OsBZR1 across different hybrid backgrounds (Figure S18).

To assess the impact of genomic variants in OsBZR1 on heterosis of field traits, we compared heterosis of the seven traits between F₁ hybrids derived from ILs carrying different haplotypes (Figure 5D–J). Except for seed setting rate and plant height, heterosis of haplotype_i/haplotype_j heterozygous F₁ hybrids did not significantly differ from the average of their homozygous counterparts (haplotype_i and haplotype_j). Strikingly, F₁ hybrids with the haplotype_i/haplotype_JR2 exhibited significantly stronger heterosis than the haplotype_i homozygotes for most traits, including yield, tiller number and plant height. In contrast, hybrids combining haplotype_j and haplotype_JR2 showed no significant yield heterosis advantage over haplotype_j homozygotes, highlighting the importance of the haplotype_i/haplotype_JR2 heterozygous state for robust yield heterosis.

2.6. Structural Variants Alter Gene Expression to Explain Variance in Heterosis

To further explore the impacts of SVs on gene expression, we performed differential expression analyses on seedling samples of the 12 ILs (Figure 6A). The differentially expressed genes (DEGs) clustered the ILs into two major groups, with C418 and R4115 in the indica group, inferring changes in genetic information from the genomic to transcriptomic levels (Figure S19). Drawing on previously reported inheritance patterns of metabolites in rice F₁ hybrids (Dan et al. 2024), we demonstrated that transcriptomic additive and partially dominant effects played more important roles than dominant effects in heterosis across the 17 traits, reinforcing the contributions of the two main effects to heterosis across different genetic levels (Figures 6B and S19).

FIGURE 6.

Relationships among genomic SVs, gene expression and heterosis. (A) Transcriptomic profiles of 12 inbred lines (ILs). Gene expression data were obtained from 15‐day‐old seedlings. (B) Changes in R ² (adjusted) for heterosis of 17 traits. Different inheritance patterns, including dominant (female and male), additive and partially dominant (0.3 and 0.7) effects, of expressed genes were used as predictive variables in partial least square (PLS) analysis on better‐parent heterosis (BPH) of 17 agronomic traits. The best regression model (with highest adjusted R ²) was selected as the final predictive model for heterosis of each trait. (C) Genomic structures of two inversions, one absence variant (AV) and five translocations (trans) across 12 ILs. (D–G) Heatmaps and PCA plots of 12 ILs based on expression levels of genes within inversion_6 and AV_6. p values indicate PERMANOVA. (H) Comparisons of R ² (adjusted) for predicting BPH of 17 traits with differentially expressed genes (DEGs) of 12 ILs, genes in inversion_6, genes in inversion_8, genes in AV_6, genes in translocation_4 and genes in translocation_12.2. (I) Changes in R ² (adjusted) of PLS models with varied number of latent factors. The additive expression levels of DEGs from 12 ILs were used as variables to predict the number of parental genomic variants. The genomic variants include single nucleotide polymorphism (SNP), insertion and deletion (InDel) and structural variant (SV). (J) R ² (adjusted) for different types of SVs. (K) R ² (adjusted) for different types of translocations. (L) Associations between SVs and heterosis. Genes linked to intronic_4, splicing_6 and upstream_7 were used to perform Spearman correlation analysis with heterosis. Their positive and negative correlations with heterosis are indicated in purple and green, respectively. Analysis of variance (ANOVA) with LSD in post hoc test (p value < 0.05) was performed in (B, H and I). Abbreviations: ESPH, elongation stage plant height; ESTN, elongation stage tiller number; FSPH, flowering stage plant height; FSTN, flowering stage tiller number; GNP, grain number per panicle; HD, heading date; MSPH, maturation stage plant height; MSTN, maturation stage tiller number; PBN, primary branch number; PL, panicle length; SBN, secondary branch number; SSPH, seedling stage plant height; SSR, seed setting rate; SSTN, seedling stage tiller number; TGW, thousand‐grain weight; TPP, tiller number per plant; YPP, Yield per plant.

We next focused on expression levels of genes within the two large inversions: inversion_6 (304 genes) and inversion_8 (331 genes; Figure 6C). Consistent with the presence/absence pattern of inversion_6, the 12 ILs were accurately classified into indica and japonica groups based on expression levels of these genes (Figure 6D,E). Unexpectedly, genes within inversion_6 with additive or partially dominant effects were predictive of heterosis in partial least squares (PLS) models (Figure S20). Based on genes in inversion_8, JR2 was closer to indica than to japonica (Figure S21). Though only a limited number of genes differed in expression levels between the indica and japonica, the genes within inversion_8 effectively explained the variance in heterosis (Figure S21). Among the three large size AVs, genes in AV_8 (22 genes) and AV_11 (22 genes) were seldom expressed in seedlings (Figure S22). For AV_6 (34 genes), most genes were expressed and effectively divided these ILs into two groups, with C418 in the indica group (Figure 6F,G). The genes in AV_6 also demonstrated predictability for heterosis comparable to the genome‐wide DEGs of 12 ILs and the two inversions (Figures 6H and S22). Our results confirmed the direct impact of SVs on gene expression, which in turn explained the variance in heterosis.

To comprehensively investigate associations between different types of parental genomic variants and expression changes in F₁ hybrids, we performed PLS analysis between additive expression levels of 1759 DEGs from the 12 ILs and the number of genomic variants in each pair of ILs. Based on the R ² (adjusted) values of PLS models with varying numbers of latent factors, the predictability of these DEGs was higher for SVs than for SNPs and InDels (Figure 6I). When SNPs and InDels were further subdivided by location or function, the predictability remained low, except for variants located in multiple regions with their number below ten (Figures S23 and S24). In contrast, the predictability varied considerably among different types of SVs, with translocation predicted at the highest accuracy (Figure 6J). Among the top five largest translocations, genes associated with translocation_8.1, translocation_12.1 and translocation_8.2 were rarely expressed in seedling samples (Figure S25). Translocation_4 was an inverted translocation on chromosome 4, detected between six indica and four japonica, except for JR2 and Balilla (Figure 6C). This translocation contained 17 genes, eight of which had detectable expression levels (Figure S26). Due to the loss of this translocation, Balilla was in indica while JR2 remained in the japonica group, suggesting a divergent effect of the same translocation on gene expression under different genetic backgrounds. The predictability of genes in translocation_4 for heterosis was lower than that of DEGs, the two inversions, and the AV_6 (Figure 6H), which was mainly caused by the limited number of expressed genes. Translocation_12.2, between chromosomes 12 and 11, was present fully in some ILs and partially in others. Based on the normalised expression levels of its 17 genes, 110 080 and Wuyunjing8 with partial translocations were clustered with indica ILs, while three indica with translocations remained in their group (Figure S26), inferring differing translocation effects on gene expression between subspecies. Despite its smaller size, the predictability of expressed genes in translocation_12.2 for heterosis was as effective as that of those within inversions and AVs (Figure 6H).

When different types of SVs (with average variant numbers over ten) were detailly analysed by location, the predictability was higher for the exonic region of AV (Figure S27), the UTR5 region of CNV (Figure S28), the splicing region of PV (Figure S29) and the intergenic region of inversion (Figure S30), suggesting diversified functions of SVs on gene expression through different locations. The predictability for translocation also varied across locations, with splicing, upstream and intronic regions having higher accuracy (Figure 6K). To further investigate the functions of different types of translocations, we selected three top size ones in combination with the identified DEGs: splicing_6, upstream_7 and intronic_4 (Figure 6C), which had available gene expression levels. These three representative translocations were invariant in indica but variable in japonica. Using a haplotype analysis strategy similar to those for OsMADS1 and OsBZR1, we classified the 12 ILs into two haplotypes based on the presence/absence of each translocation. Then, we compared the heterosis of the 17 agronomic traits between hybrids with heterozygous and homozygous haplotypes to investigate their relationships with heterosis. The associations varied across translocations, with the associated genes showing either positive or negative correlations with heterosis (Figure 6L). We also analysed three other representative translocations, including splicing_11, upstream_1 and intronic_11, which were detected in at least five ILs but with nearly undetectable gene expression (Figure S25). These translocations showed distinct associations with heterosis (Figure 6L). After the haplotype analysis was extended to all the above‐mentioned typical SVs, we found the largest number of associations for inversion_6 and AV_6 (n = 10; Figure 6L). Then, the upstream_7, splicing_11 and upstream_1 were associated with the heterosis of nine traits. The upstream_7 was the only translocation for yield heterosis, warranting further validation of this region or its associated gene by sequence modulation technologies. Taken together, the heterozygous haplotypes of SVs displayed significant heterosis to their homozygous ones, supporting the overdominance model in rice heterosis.

3. Discussion

The genetic mechanism of heterosis has been a central debate in plant genetics for over a century. The overdominance model emphasises the contribution of genomic heterozygosity and the co‐existence of nondefective alleles (East 1909, 1936; Shull 1908, 1911). In this study, genomic variants in parental lines are at heterozygous states in F₁ hybrids. SNPs, InDels, PAVs, CNVs, inversions and translocations represent heterozygosity of single bases (1 bp), several bases (2–50 bp), some bases (> 50 bp), sequence repeat number, sequence direction and sequence location, respectively (Figure S31). When genomic variants are combined for haplotype analysis, F₁ hybrids with heterozygous haplotypes show heterosis to those with homozygous haplotypes, underscoring the critical role of genomic heterozygosity in rice heterosis. For the S5‐ORF5 locus, typical indica‐japonica F₁ hybrids with heterozygous ORF5+ and ORF5−, which are functional alleles of aspartic protease (Chen et al. 2008; Yang et al. 2012), exhibit negative heterosis for seed setting rate. In contrast, hybrids carrying the ORF5n, which is assigned as a nonfunctional allele due to a 94 bp AV spanning the translation start site, alleviate the negative heterosis. Thus, the ‘nonfunctional’ ORF5n allele shows significant function to heterosis for seed setting rate. Since inbred lines carrying any of the three alleles have normal embryo‐sac fertility and seed setting rate, it is not appropriate to classify ORF5n as a dominant superior allele and ORF5+/ORF5− as recessive detrimental alleles under the dominance model. Similarly, heterozygous haplotypes of OsBZR1 (haplotype_i/haplotype_JR2) outperformed homozygous ones for yield and multiple yield‐related traits. Most importantly, our haplotype‐based analysis of representative translocations and other SVs consistently indicates that the heterozygous state confers heterosis over the homozygous counterparts.

The overdominance model also posits the relationship between parental genetic disparity and the degree of heterosis. Previous studies using low‐ or high‐density DNA markers have supported the close associations between genetic distance or genetic variants and heterosis of F₁ hybrids in rice (Dan et al. 2014; Huang et al. 2015; Xiao et al. 1996; Zhang et al. 1996). We here offer a more comprehensive genomic variant picture based on high‐quality genome assemblies and demonstrate that the number of genomic variants in parents is a powerful predictor of heterosis for multiple agronomic traits. Crucially, the relationships differ between inter‐ and intrasubspecific hybrids, especially for heterosis of seed setting rate, tiller number per plant, and yield per plant. Based on the genome‐wide heterozygosity of genomic variants, particularly SVs, and their correlations with heterosis, we support the overdominance model and propose that heterozygosity of alleles/sequences underlies heterosis in rice.

Although we generated reference‐level genome assemblies, telomere‐to‐telomere (T2T) genomes for all investigated inbred lines are needed to display the complete genomic landscape in subsequent heterosis studies. Concurrently, advanced variant detection and description tools should be developed to accurately characterise genomic variants in parental inbred lines. While large hybrid populations (often hundreds of F₁ hybrids) are standard for robust heterosis studies, and hybridization and phenotyping are feasible in rice, generating hundreds of reference‐ or T2T‐level genomes remains challenging for most research teams currently. Complete or half diallel‐cross designs, which efficiently utilise parental genomic information, are practical strategies for deeper genomic investigation on rice heterosis. For parental lines, only five inbred lines used in this study are parents of elite hybrid rice varieties. F₁ hybrids randomly derived from these and other lines do not show significant heterosis for traits like panicle length and branch number. Crosses between those proven, elite parents reliably produce heterotic hybrids and enable a more precise dissection of heterosis. Additionally, the representativeness of parental inbred lines is important to outcomes. The 12 inbred lines used here capture only a portion of rice genetic diversity. The inclusion of accessions from aus, tropical japonica, basmati and wild accessions (e.g., rufipogon) will provide updated genomic insights into rice heterosis.

The SVs detected in our study ranged from about 57–168 Mb, constituting approximately 14%–42% of the genomes. Consistent with previously reported patterns of genomic variation (Qin et al. 2021; Zhang, Fu, et al. 2022), PAVs constitute the majority of SVs (74.0%). These SVs, particularly subspecies‐specific inversions and AVs: inversion_6 and AV_6, provide a structural basis for indica–japonica divergence and genetic variation among inbred lines. Inversion_6, which has been documented across multiple indica (e.g., 9311, Xiangwanxian12, Kasalath, Fengyuezhan, Huazhan, Tianfeng) and japonica (e.g., Nipponbare, Wuyunjing7, Xiushui134, Chugeng37) genome assemblies (Wei et al. 2024; Zhang, Wang, et al. 2022), represents a key contributor to rice subspeciation. Importantly, specific SVs are associated with heterosis of key agronomic traits, highlighting the application potential of SVs in genomic prediction. As noted, the 94 bp AV in S5‐ORF5 and the complex PAVs linked to OsBZR1 drive heterosis for seed setting rate and yield, respectively. We propose that SVs distinguishing indica from japonica or divergent haplotypes are candidates for precise engineering to harness strong intersubspecific heterosis. The identified translocations, particularly intronic_4, splicing_6 and upstream_7 related to changes in gene expression levels, are important targets. Although precise modulation of genomic SVs remains challenging, we believe that advances in genome editing technologies, such as DualPE (Zhao et al. 2025), will elucidate the function of SVs and facilitate the utilisation of heterosis in crops. Prior to establishing precise editing methods for SVs, for example, translocations, the effects of specific SVs on heterosis can be indirectly validated via population genetic strategies. One strategy involves constructing near‐isogenic lines or single‐segment substitution lines for individual SVs using DNA markers. Phenotypic comparisons between the recipient parent and these near‐isogenic lines or single‐segment substitution lines can then provide preliminary assessments of target SVs. After these lines were crossed to a common inbred line, the contribution of investigated SVs can be further evaluated by comparing heterosis in F₁ hybrids with heterozygous versus homozygous states of the target SVs. An alternative strategy employs recombinant inbred lines. Firstly, two inbred lines carrying target SVs are crossed to generate heterozygous F₁ hybrids and subsequent recombinant inbred lines. These recombinant inbred lines—typically requiring at least five generations of self‐crossing to achieve phenotypic uniformity—are then crossed with a male‐sterile line to produce F₁ hybrids for heterosis quantification of corresponding traits. To classify genotypes/haplotypes, the recombinant inbred lines and the sterile line are analysed using standard PCR assays targeting the SVs. Finally, the functional impact of single SVs can be assessed by comparing heterosis between heterozygous and homozygous genotypes/haplotypes.

Genomic SVs always induce gene expression changes (Alonge et al. 2020; Shang et al. 2022). Moving beyond individual loci, our transcriptomic analyses reveal that rice genomic SVs, especially large‐size inversions and AVs, not only correctly classify inbred lines by subspecies but also effectively explain the variance in heterosis through the cumulative, additive expression of multiple genes. These SVs are disproportionately important drivers of transcriptional changes, with translocations (e.g., translocations located in splicing, upstream and intronic regions) outperforming others. We detect associations among SVs, expression changes and heterosis based on seedling transcriptomes. However, many genes are expressed in specific tissues (e.g., young panicles), stages or environments. In addition to gene expression, genomic SVs may be linked to DNA methylation levels, chromatin accessibility (Wang, Xi, et al. 2024) and histone modifications (Qi et al. 2024). Therefore, the collection of diversified genomic‐downstream data, including epigenomic, transcriptomic and translational data, is essential to design optimal SV heterozygosity in generating rice heterosis.

4. Materials and Methods

4.1. Plant Materials

Twelve rice inbred lines (ILs) were selected from accessions reported in previous studies (Dan et al. 2020, 2024, 2016, 2014). Based on clustering results from metabolite profiles generated using GC‐ or LC–MS platforms (Dan et al. 2020, 2016), low‐density SSR markers (Dan et al. 2014) and high‐density SNPs or InDels (Dan et al. 2024), these 12 inbred lines are representatives of indica and japonica accessions. Plants of the selected inbred lines for genome sequencing were grown under long‐day conditions, maintaining a 16‐h light/8‐h dark photoperiod at 30°C in a plant culture room. The seedlings were sown in 96‐well PCR plates (without bottoms) and placed in a plastic basin filled with Yoshida nutrient solution (pH = 5.5). After 20 days of growth, leaves from each genotype were harvested for experimentation. For full‐length RNA sequencing of YB and Balilla, the plants were divided into four tissue parts: root, middle leaf (the youngest leaf), leaf without middle leaf and stem.

Additionally, seedlings of the 12 rice inbred lines were placed in an intelligent breeding room for RNA sequencing. The environmental conditions were set to 28°C with a 16‐h light/8‐h dark photoperiod and controlled humidity. Each inbred line occupied half of a PCR plate and was maintained in a single basin containing four litres of nutrient solution, which was changed every 3 days. Fifteen‐day‐old seedlings (without roots) were harvested and frozen in liquid nitrogen. The middle five plants from each row were pooled to create one biological replicate, with three biological replicates established for each genotype.

The 12 inbred lines were crossed using a complete diallel design. As previously detailed (Dan et al. 2024), a total of 17 agronomic traits were recorded for both the inbred lines and their corresponding F₁ hybrids. These traits included grain yield per plant (YPP), grain number per panicle (GNP), seed setting rate (SSR), thousand‐grain weight (TGW), plant height/tiller number (at seedling, elongation, flowering and maturation stages) (SSPH, ESPH, FSPH, MSPH and SSTN, ESTN, FSTN, MSTN), heading date (HD), tiller number per plant (TPP), panicle length (PL) and the number of primary and secondary branches per panicle (PBN and SBN). Better‐parent heterosis was calculated using the equation: BPH = (F₁—P _high)/P _high, where P _high represents the value of the high parent. For the heading date, parents with shorter heading dates were designated as P _high, and the calculated values were subsequently multiplied by −1.

4.2. Genomic Sequencing

SMRTbell libraries (~17 kb) for each rice inbred line were constructed using the SMRTbell Express Template Prep Kit 3.0 and sequenced on the Pacbio Revio platform (Pacific Biosciences) at Frasergen Bioinformatics (Wuhan, China). The SMRTbell libraries were size‐selected using the BluePippin system (Sage Science) and purified with a 1× concentration of SMRTbell cleanup beads. Library quality was assessed using Femto pulse and Qubit 3.0 (Life Technologies), resulting in an average of 12.31 Gb of HiFi reads (approximately 30.46× coverage of the genome) obtained for each inbred line.

Hi‐C libraries for YB and Balilla were prepared and sequenced using the BGI PE150 platform, generating a total of 51.81 Gb and 49.27 Gb of reads (approximately 127.80× coverage of the genome) for the two samples, respectively. The raw data were filtered using Trimmomatic (v0.40; Bolger et al. 2014) with parameters LEADING:3 TRAILING:3 SLIDINGWINDOW:4:15 MINLEN:15. The quality of the clean data was evaluated with default parameters using FastQC (Wingett and Andrews 2018).

4.3. RNA Sequencing

Full‐length RNA sequencing technology was employed to aid in the annotation of genes in YB and Balilla. Total RNA was extracted from four tissues of the seedling samples, and equal amounts of RNA were pooled before constructing SMRTbell libraries. Full‐length cDNA was synthesised using the SMARTer PCR cDNA Synthesis Kit (Clonetech) with the primer sequences ‘AAGCAGTGGTATCAACGCAGAGTACATGGGG’ and ‘AAGCAGTGGTATCAACGCAGAGTAC’. The products were purified and size‐filtered (> 1 kb) prior to sequencing. The PacBio Sequel II platform at Frasergen Bioinformatics was used for sequence analysis, with raw data (an average of 36.14 Gb) filtered and preprocessed using SMRTlink (v10.1, ‐‐minLength = 50). Iso‐seq analysis was performed to obtain full‐length and non‐concatemer (FLNC) sequences. These FLNC sequences were mapped to reference genomes using GMAP (V19.06.10; Wu and Watanabe 2005), retaining those with the highest PID to determine gene structure.

To investigate the impact of structural variants on gene expression changes, RNA sequencing was also conducted on seedlings of the 12 rice inbred lines (with three biological replicates per genotype). Total RNA was extracted for each sample to construct a DNA library. The Illumina Hiseq PE150 platform at Frasergen Bioinformatics was selected to generate 150 bp paired‐end reads, with more than 6 Gb of raw data per sample. The raw data were filtered using SOAPnuke (v2.1.0, ‐lowQual = 20, ‐nRate = 0.005, ‐qualRate = 0.5; Chen et al. 2018). Clean reads were aligned to the YB genome using Hisat2 (v2.2.1, ‐‐no‐unal; Kim et al. 2019). Transcripts were assembled using StringTie (Pertea et al. 2015), and clean reads were mapped to reference transcripts using bowtie2 (v2.3.5, ‐‐sensitive ‐k 50 ‐‐no‐mixed ‐‐no‐discordant; Langmead 2010). The read count of each transcript was calculated using RSEM (v1.3.3; Li and Dewey 2011).

4.4. Genome Assembly

Draft genomes of the 12 rice inbred lines were assembled using HiFi reads with parameter ‐l3 via HiFiasm (v0.19.5; Cheng et al. 2022). The clean Hi‐C reads for YB and Balilla were aligned to contigs using Juicer (v1.6; Durand et al. 2016). Contact matrices were generated with 3D‐DNA (v180922, ‐r0; Dudchenko et al. 2017), and corrections were applied using JuiceBox (v1.11.08). The genomes of the remaining 10 ILs were assembled based on the genomes of YB and Balilla using Ragtag (v2.1.0, scaffold; Alonge et al. 2022). We assessed the accuracy, identity, and completeness of the 12 assemblies. The completeness of the assemblies was assessed using Benchmarking Universal Single‐Copy Orthologue (v3.0.2, embryophyta_odb10) analysis (Simao et al. 2015). The long terminal repeats (LTR) assembly index (LAI) was calculated for each chromosome using LTR tools (Ou et al. 2018), including LTR_retriever (v2.9.0, ‐infinder ltrfinder.scn ‐inharvest ltrharvest.scn), LTR_FINDER (v1.0.7, ‐D 15000 ‐L 7000 ‐C ‐M 0.7) and GenomeTools (v1.6.2, ltrharvest ‐similar 70 ‐vic 10 ‐seed 20 ‐seqids yes ‐maxlenltr 7000 ‐maxtsd 6). HiFi reads were re‐mapped to the assemblies with parameters ‘‐axe map‐hifi’ using Minimap2 (v2.21; Li 2018). Additionally, consensus quality values and completeness of the assemblies were scored using a k‐mer‐based method with the Merqury pipeline (v1.3; Rhie et al. 2020).

4.5. Genome Annotation

Repetitive sequences (including tandem repeats and transposable elements), protein‐coding genes and noncoding RNAs were annotated for the genome assemblies of YB and Balilla. Tandem repeats were identified using the Tandem Repeats Finder (TRF, v4.09.1, 2 7 7 80 10 502 000 ‐d ‐h; Benson 1999). Transposable elements were annotated through an integrated approach combining de novo and homology‐based (Repbase, www.girinst.org/repbase (Bao et al. 2015)) strategies. The de novo repeat library was constructed using LTR_FINDER (v1.0.7, ‐w 2 –C; Xu and Wang 2007) and RepeatModeler (v2.0.1; Flynn et al. 2020). Transposable elements were detected using RepeatMasker (v4.1.2; Tarailo‐Graovac and Chen 2009) with parameters ‘‐nolow ‐no_is ‐norna ‐parallel 2’, searching against both the Repbase (Bao et al. 2015) and de novo repeat library. Additionally, RepeatProteinMask (1.36) (Tarailo‐Graovac and Chen 2009) with parameters ‘‐engine ncbi ‐noLowSimple ‐pvalue 0.0001’ was utilised to identify transposable elements in Repbase (Bao et al. 2015).

To annotate gene structure and function, we integrated three strategies: homology‐based (using five genomes), de novo and transcriptome‐based (full‐length RNA sequencing) predictions. In the homology‐based approach, protein‐coding sequences from the genomes of 9311 (Yu et al. 2002), Minghui63 (Zhang et al. 2016), Zhenshan97 (Zhang et al. 2016), R498 (Du et al. 2017) and Nipponbare (Shang et al. 2023) were compared with the YB and Balilla genomes using tblastn (v2.11.0+, −evalue 1e‐05; Gertz et al. 2006). Gene structures of the aligned sequences were predicted using Exonerate (v2.4.0, ‐‐model protein2genome; Slater and Birney 2005). De novo annotation was conducted using Augustus (v3.4.0, ‐‐noInFrameStop = true ‐‐strand = both; Stanke et al. 2006), GlimmerHMM (v3.0.4; Majoros et al. 2004), and GeneMark (v4.65; Borodovsky and McIninch 1993). The annotated gene sets were integrated to nonredundant sets using MAKER (v3.01.03; Holt and Yandell 2011), and further polished by incorporating full‐length RNA sequencing results using PASA (v2.4.1; Haas et al. 2003). RNA‐sequencing results were aligned using HiSat2 (v.2.2.1; Kim et al. 2019), and transcripts were assembled with StringTie (v.2.1.7; Kovaka et al. 2019). Additionally, transcripts were de novo assembled using Trinity (v2.8.5; Grabherr et al. 2011). Gene functional analysis was conducted with Diamond (v2.0.7, ‐c 1 ‐f 5 ‐e 1e‐5 ‐‐very‐sensitive ‐k 10; Buchfink et al. 2021) against five databases: Nonredundant (NR, ftp://ftp.ncbi.nlm.nih.gov/blast/db/FASTA/nr.gz), Kyoto Encyclopedia of Genes and Genomes (KEGG) (Kanehisa et al. 2014), Gene Ontology (GO) (Ashburner et al. 2000), TrEMBL (Boeckmann et al. 2003) and Swiss‐Prot (Boeckmann et al. 2003). Protein domains were annotated using InterProScan (v5.50‐84.0, ‐‐applications Pfam; Jones et al. 2014) based on the InterPro database (Mitchell et al. 2015).

Non‐coding RNAs, including rRNA, tRNA, snRNA and miRNA, were also annotated for the two genomes. tRNA was identified using tRNAscan‐SE (v2.0.9) (Lowe and Eddy 1997), while rRNA was detected using RNAmmer (v1.2, ‐S euk ‐multi ‐m lsu,ssu,tsu; Lagesen et al. 2007). miRNA and snRNA were annotated with Infernal (v1.1.4; Nawrocki and Eddy 2013) using Rfam (v14.6; Griffiths‐Jones et al. 2005). miRNA and rRNA were present in higher percentages within the genomes. The completeness of the genomic annotations was further assessed using BUSCO (v5.2.2, embryophyta_odb10; Simao et al. 2015). Here, 98.5% and 98.6% of the gene sets in the two genomes contained complete BUSCOs in embryophyta.

For the annotation of the remaining 10 genomes, Liftoff (Shumate and Salzberg 2021) was employed. The YB genome was chosen as the reference genome for five indica accessions: R465, Mianhui725, 9311K, 610234 and Qianlijing. The Balilla genome served as the reference for five japonica accessions: R4115, C418, 110080, Wuyunjing8 and JR2.

4.6. Pan‐Genome Analysis

Protein‐coding genes of the 12 inbred lines were initially aligned using BLASTp (‐evalue 1e‐5; Altschul et al. 1997). The BLAST results were processed with OrthoFinder (v2.3.12; Emms and Kelly 2019) to classify gene families. Gene families shared among all 12 inbred lines were defined as core‐gene families, while those identified in more than 90% of inbred lines were termed softcore gene families. Gene families specific to a single inbred line were classified as private gene families, and the remaining gene families were considered dispensable. Genes in the four types of gene families were further analysed in the Pfam (Mistry et al. 2021), GO (Ashburner et al. 2000), and KEGG (Kanehisa et al. 2014) databases to determine their conservation.

4.7. Identification of Genomic Variants

Genomic variants, including single nucleotide polymorphisms (SNPs), insertions and deletions (InDels) (2–50 bp) and structural variants (SVs), were detected in the 12 rice inbred lines, with the genome of YB selected as the reference genome and the remaining 11 as query genomes. SNPs and InDels were detected using MUMmer4 (nucmer: ‐c 90 ‐l 40, delta‐filter: −1, show‐SNPs: ‐ClrTH) (Marcais et al. 2018). SVs, including presence–absence variants (PAVs), inversions, translocations and copy number variants (CNVs), were detected with the assistance of SyRI (v1.4, nucmer: ‐c 90 ‐l 40 ‐t 16 ref.fa query.fa, delta‐filter: ‐1 ‐i 90 ‐l 100 ref_query.delta, show‐coords: ‐THrd; Goel et al. 2019). The distribution of SNPs, InDels, and SVs was analysed across various genomic regions or functions, including: (1) Upstream (1 kb regions before the gene start codon); (2) Downstream (1 kb regions after the gene stop codon); (3) Intergenic regions; (4) UTR3 (3′ untranslated regions); (5) UTR5 (5′ untranslated regions); (6) Splicing junction regions (< 10 bp); (7) Intronic regions; (8) Exonic regions; (9) Synonymous or nonsynonymous mutants; (10) Frameshift or nonframeshift insertions/deletions; (11) Stop‐gain (gain of stop codons); (12) Stop‐loss (loss of stop codons). Variants simultaneously located in two regions, such as UTR5 and UTR3, were shown as Region 1; Region 2. The functions of SNPs, InDels and SVs were annotated with ANNOVAR (Wang et al. 2010).

4.8. Pan‐SV Analysis

A pan‐SV analysis was conducted for SVs of the 11 query genomes and the results were further used for constructing a graph‐based genome using Vg (v1.43; Garrison et al. 2018). For SVs excluding PVs, an SV was assigned when the ratio of overlapping to total locations in two inbred lines exceeded 50%. For PVs, a PV was designated when the distance between PV locations in two inbred lines was less than 5 bp and the length ratio between the shorter and longer samples was greater than 50%. Circles illustrating genomic variants on chromosomes were generated using Circos (Krzywinski et al. 2009). Genetic relatedness among the 12 inbred lines was visualised using Cytoscape (Reimand et al. 2019).

4.9. Construction of Phylogenetic Trees

Phylogenetic trees based on the number of differential SNPs, InDels, and CNVs among the 12 rice inbred lines were built using the neighbour‐joining method with MEGA (v10.2.6; Stecher et al. 2020). The number of SNPs, InDels and CNVs of the inbred lines was retrieved from a previous report (Dan et al. 2024). Population structure analysis was performed using Admixture (version 1.3.0) based on SNPs.

For phylogenetic analysis of Phr1 and the S5 loci, gene sequences from reference inbred lines were obtained from the GenBank database. The Phr1 sequences of Nipponbare and 9311 corresponded to accession numbers DQ532376 and DQ532391, respectively (Yu et al. 2008). For S5‐ORF3‐5 (Chen et al. 2008; Yang et al. 2012), the Balilla sequences were retrieved under accession numbers JX138498, JX138502, EU889294. Gene sequences from the 12 newly assembled inbred lines were extracted and aligned using Jalview (version 2.11.4.1) with default clustal parameters (Waterhouse et al. 2009). Phylogenetic trees were constructed with the neighbour‐joining method. Genotypes of inbred lines were assigned based on previous reports (Chen et al. 2008; Yang et al. 2012; Yu et al. 2008). Gene structures were visualised using the exon–intron graphic maker (http://wormweb.org/exonintron).

To investigate genomic variants of OsMADS1 and OsBZR1, especially SVs in the promoter region of OsBZR1, sequences (including 6345 bp before the ATG of OsBZR1) of the 12 ILs were used for analysis. In sequence alignment, the mafft combined with E‐INS method was chosen. The dendrogram tree of the 12 ILs based on phenotypic data of seven traits was implemented using MetaboAnalyst 6.0 (Pang et al. 2022), with Spearman and ward as the distance measure and clustering algorithm, respectively.

4.10. Phenol Reaction Experiment

Seeds from 14 inbred lines were harvested at maturity and stored at −20°C. For each inbred line, three seeds were chosen and treated with a 2% (volume/volume) phenol solution at 37°C. Changes in hull colour of the seeds were recorded after 1, 2, 3, 5, 7, 31, 62, 93 and 124 days of treatment. After 7 days of treatment, the seeds were placed at room temperature.

4.11. Normalisation and Differential Analysis of Transcriptomics Data

Read counts of 53 564 genes in each experimental sample were used for data normalisation and differential analyses through MetaboAnalyst 5.0 (Pang et al. 2022). The interquantile range was selected to filter genes, with a threshold set at 40%. Sum, none and auto‐scaling were chosen for sample normalisation, data transformation and data scaling, respectively. Differential analysis among the 12 inbred lines was implemented using the ANOVA module in MetaboAnalyst 5.0 (Pang et al. 2022), with a raw p value cutoff set at 0.001 for post hoc analysis. Principal component analysis (PCA) and heatmap results for the 12 inbred lines were obtained through corresponding modules on the website. PERMANOVA was conducted in PCA and the permutation number was 999. Euclidean for distance measure and ward for clustering method were set in heatmaps. Differential genes were selected from the original read count table and averaged across the three biological replicates. The averaged data were further filtered and normalised for each genotype with the above parameters. Due to the limited number of genes in inversions and absence variants, no filtering was applied to these genes. The genes in three large AVs, including AV_6, AV_8 and AV_11, were analysed. Despite AV_8 and AV_11 being larger in size, the genes in these AVs were rarely expressed in the seedling samples. Only genes in AV_6 were further analysed.

To investigate the relationship between the expression level of OsBZR1 (Os07g0580500) and heterosis, gene expression data and seedling length (15‐day‐old seedlings) of six pairs of reciprocal hybrids and corresponding parental inbred lines were obtained from a previous study (Dan et al. 2024). The transcriptomic data were Log2‐normalised using NetworkAnalyst 3.0 (Zhou et al. 2019). Since several genotypes had six biological replicates, seedling length of the first three replicates was used for the calculation of better‐parent heterosis and Spearman correlation analysis. To determine inheritance patterns of OsBZR1, the reciprocal hybrids from two inbred lines were treated as a single genotype. Average expression levels of OsBZR1 were calculated between each pair of inbred lines and were treated as additive expression levels. Gene expression levels were compared for the parents, F₁ hybrids, and additive expression levels. Inheritance patterns were assigned based on previous rules (Dan et al. 2024).

4.12. Partial Least Square (PLS) Analysis

PLS analysis was conducted to investigate the relationship between inheritance patterns of expressed genes and better‐parent heterosis of 17 agronomic traits using SPSS 20.0. The maximum number of latent factors was set to 50. The different inheritance patterns were manually calculated based on normalised gene expression levels using equations outlined in a previous study (Dan et al. 2024). The genetic effects included: (1) Additive effect: average levels of the female (P _f) and male (P _m) parents; (2) Partially dominant effects (PD): partially dominant _0.3 (PD_0.3) = P _m + 0.3 × (P _f−P _m) and partially dominant _0.7 (PD_0.7) = P _m + 0.7 × (P _f−P _m); (3) Dominant effects: expression levels of the female (dominant_female) and male (dominant_ male) parents. The model with the highest R ² (adjusted) was selected as the final model for each trait.

PLS was also used to study the relationship between differentially expressed genes in 12 inbred lines and the number of parental genomic variants. The additive effect of the differential genes was used as predictive variables to predict the number of unique variants between each two inbred lines.

4.13. Statistical Analysis

Boxplots were created using BoxPlotR (Spitzer et al. 2014). Analysis of variance (ANOVA) with the least significant difference for post hoc comparisons was performed using SPSS 20.0. This analysis compared the following parameters: lengths of genes, lengths of coding sequences, lengths of exons, lengths of introns, GC content of genes, GC content of coding sequences, number of exons and number of introns across seven genomes (including two newly assembled genomes, YB and Balilla, along with five representative rice genomes). Additionally, we assessed the number of genomic variants and the degrees of better‐parent heterosis for 17 traits in F₁ hybrids from both intra‐ and intersubspecies. The adjusted R ² values of partial least square models for heterosis across traits were also compared. An independent samples t‐test (weighted) was conducted to evaluate the number of correlations between unique and common genomic variants and heterosis. SPSS was also employed to conduct Spearman correlation to investigate the relationships between different types of unique genomic variants and between genomic variants and heterosis.

Author Contributions

Z.D. conceived and designed the project. Z.D. and Y.C. prepared rice seedling samples, conducted genomic, transcriptomic, phenotypic and statistical analyses, and interpreted the results. W.H. collected the rice accessions. Z.D. and W.H. constructed the diallel‐cross population and recorded the phenotypes. Z.D., Y.C. and W.H. wrote the manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (31801439 to Z.D., 32472185 to W.H. and 32101667 to Y.C.), the China Postdoctoral Science Foundation (2019M660186 to Z.D., 2022T150500 to Z.D. and 2023T160497 to Y.C.), the Key Research and Development Program of Hubei Province (2022BFE003 to W.H.), the National Key R&D Program of China (2017YFD0100400 to W.H.) and the Hubei Agriculture Science and Technology Innovation Center Program to W.H.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Cross‐validation error and population structure of 12 rice inbred lines.

Figure S2: Pedigree profiles of six inbred lines used in this study.

Figure S3: Heterosis of 16 agronomic traits in F₁ hybrids from intra‐ or intersubspecies.

Figure S4: Contact matrices and collinearity diagrams for genomes of YB and Balilla.

Figure S5: Long terminal repeat assembly index (LAI) for 12 chromosomes of 12 rice accessions.

Figure S6: Comparison of two newly assembled genomes with five established rice genomes.

Figure S7: Genomic comparisons between Mianhui725, 610 234 and YB.

Figure S8: Genomic comparisons between 9311K, R465 and YB.

Figure S9: Genomic comparisons between Qianlijing, C418 and YB.

Figure S10: Genomic comparisons between R4115, Wuyunjing8 and YB.

Figure S11: Genomic comparisons between 110 080, JR2 and YB.

Figure S12: Genomic variants displayed on the genome of YB.

Figure S13: Genetic relatedness of 12 rice inbred lines based on different types of genomic variants.

Figure S14: Number of genomic variants for F₁ hybrids from intra‐ and intersubspecies.

Figure S15: Correlations between the number of structural variants and heterosis of intra‐ and intersubspecific F₁ hybrids.

Figure S16: Contributions of genomic variants in OsMADS1 to heterosis.

Figure S17: Dendrogram of 12 inbred lines based on phenotypic data of seven agronomic traits.

Figure S18: Expression levels of OsBZR1 in inbred lines and F₁ hybrids.

Figure S19: Transcriptomic additive and partially dominant effects contribute to heterosis of 17 traits.

Figure S20: Associations between expression levels of genes in inversion_6 and heterosis.

Figure S21: Associations between expression levels of genes in inversion_8 and heterosis.

Figure S22: Genomic structure of two AVs and the prediction of heterosis with AV_6.

Figure S23: Adjusted R ² for different types of SNPs.

Figure S24: Adjusted R ² for different types of InDels.

Figure S25: Genomic structure and expression patterns of five translocations.

Figure S26: Associations between genes in translocation_4/translocation_12.2 and heterosis.

Figure S27: Adjusted R ² for different types of AVs.

Figure S28: Adjusted R ² for different types of CNVs.

Figure S29: Adjusted R ² for different types of PVs.

Figure S30: Adjusted R ² for different types of inversions.

Figure S31: Heterozygosity of genomic variants in F₁ hybrids.

Table S1: Genome assembly profile of 12 rice inbred lines.

Table S2: Consensus quality values and completeness of the 12 assemblies.

Table S3: Hi‐C profile of two rice genomes.

Table S4: Annotation of two rice genomes.

Table S5: Genomic variants of each pair of inbred lines and heterosis of their F1 hybrids.

Data Availability Statement

Raw reads of PacBio SMRT sequencing and RNA sequencing of the 12 rice inbred lines have been deposited in the NCBI database under BioProject accession PRJNA1178436. Genome assemblies of the 12 rice inbred lines are available from Figshare repository with DOI: https://doi.org/10.6084/m9.figshare.28397960. Source data are available with this paper.