Structure and function of rice hybrid genomes reveal genetic basis and optimal performance of heterosis

Heterosis or hybrid vigor describes superior performance of F1 hybrids over parental inbred lines and is a cornerstone of modern crop improvement. Despite historical successes in maize and rice, the quantitative genomic basis of heterosis and practical tools for predicting optimal hybrid crosses remain limited. Three non-mutually exclusive hypotheses—dominance (genetic complementation), overdominance (heterozygote advantage) and epistasis—have been proposed and supported to varying extents across species. In rice, numerous heterosis-related QTLs have been reported, and evidence suggests nonlinear effects of multiple heterozygous loci and dosage sensitivity underlie hybrid performance. Yet, a comprehensive genome-wide analysis capturing breeding footprints across decades of hybrid breeding, coupled with a broadly applicable prediction framework, has been lacking. This study resequences and phenotypes a nationwide panel of commercial rice hybrids and segregating progenies to (1) quantify phenotypic changes and genomic diversity across breeding eras, (2) identify and characterize heterotic loci and their dominance properties, (3) dissect the genetic basis of intersubspecific (indica-japonica) heterosis, and (4) develop and validate a genomic selection model to optimize hybrid combination design.

Classical models of heterosis include dominance (complementation of deleterious recessives), overdominance (heterozygote advantage at single loci), and epistasis (interactions between loci). Empirical support exists across crops: single-locus overdominance/pseudo-overdominance in tomato and sorghum, and genetic complementation contributing to heterosis in Arabidopsis, maize, and rice. Prior genomic analyses in rice identified numerous QTLs for heterosis and suggested partial dominance of heterozygotes and dosage sensitivity. Introgression and divergent selection have been shown to shape hybrid genome architecture. Key genes affecting rice yield and quality (for example, NAL1, Waxy, DEP1, Ghd7/8, Hd1, GS3, GW5/7) have been implicated in agronomic performance and adaptation, providing candidates for assessing breeding footprints and heterosis mechanisms.

Materials and phenotyping: The study assembled 2,839 commercial rice hybrids released from 1976–2020 from the China National Rice Research Institute. Hybrids were grouped into three breeding periods: Y1 (1976–2000), Y2 (2001–2010), Y3 (2011–2020). Five cytoplasm types were considered: Wild-abortive (WA), Boro II (BT), Honglian (HL), Twoline-Jap (TJ) and Twoline-Ind (TI), with HL excluded from downstream analyses due to small size. Eighteen representative hybrids (10 indica-indica and 8 indica-japonica; 5 WA, 3 TJ, 3 TI, 7 BT) generated 9,839 F2 individuals (18 populations; sizes 340–949). All F1 hybrids (2020, Hangzhou) and F2 individuals were phenotyped for heading date, morphology (flag leaf length/width, plant height, tiller angle), yield components (full/total grain number per plant, valid panicle number, full grain number per panicle, kilo-grain weight, seed setting rate, yield per plant), and grain quality (amylose content, gel consistency, chalkiness, chalky grain percentage, grain translucency, grain shape) using standardized protocols. A total of 123 parent–hybrid trios enabled better-parent heterosis assessment. Sequencing and variant discovery: Hybrid genomes were resequenced to ~35× depth using NovaSeq 150 bp paired-end reads. Reads were quality-filtered (Trimmomatic) and aligned to Oryza sativa Nipponbare IRGSP 1.0 (BWA), with variant calling via GATK (HaplotypeCaller/GenotypeGVCFs) and hard filtering. The dataset yielded 5,222,902 high-quality SNPs and 1,701,091 indels (≤50 bp). Structural variants (≥51 bp) were detected (graph-based genome) in 964 representative hybrids, yielding 22,555 SVs. Low-coverage (~0.2×) WGS was performed for F2s; genotypes were inferred by binning against dense parental SNPs and imputed. Population and diversity analyses: Nucleotide diversity (VCFtools), PCA (GCTA) and ADMIXTURE (K=2–5) were performed (using 4DTv sites). Kinship (EMMAX) and visualization (Cytoscape) assessed relatedness within periods. Japonica introgression was mapped using 830,245 indica–japonica differential SNPs (from 19 indica, 23 temperate japonica accessions). A 199-SNP sliding window (1-SNP step) labeled homozygous japonica or heterozygous indica/japonica introgression when ≥120 loci met criteria; adjacent segments were merged. Method accuracy was validated on Nipponbare (japonica) and Shuhui498 (indica). GWAS and effect quantification: GWAS was conducted with MLM (TASSEL) for three cohorts: 2,716 indica-indica hybrids; 4,497 indica-indica F2s; 5,342 indica-japonica F2s. SNP filters included missing rate ≤10% and MAF ≥5% (PLINK) for hybrids; imputed high-density SNPs were used for F2s (1,296,386 in indica-indica; 1,059,427 in indica-japonica). The top two PCs and a kinship matrix were included as covariates. Significance threshold was 1×10^-6. Dominance-to-additive ratio (d/a) was computed from genotype effects of peak SNPs: a = (A−C)/2, d = M − (A+C)/2. For traits where lower values are favorable (for example, chalkiness), −d/a was used. Loci with <5 observations for any homozygote were excluded. QTL mapping in F2s (IciMapping) provided independent d/a estimates. Phenotypic variance explained (PVE) by 1 Mb windows around association peaks was estimated with mixed models including additive and dominance relationship matrices (sommer). Breeding-favorable genotypes were defined by prior QTNs where available, or by effect directions consistent with agronomic targets (for example, larger leaves, higher yield, lower chalkiness), with heading date favoring slightly shorter durations. Genomic selection model: A GBLUP model (sommer) using 88,909 SNPs and phenotypes for seven traits (yield per plant, valid panicle number, full grain number per panicle, seed setting rate, plant height, heading date, grain shape) was trained on 12,678 individuals (2,839 hybrids + 9,839 F2s). Models included additive (A) and additive+dominance (A+D) formulations; six-fold cross-validation (three iterations) assessed accuracy via Pearson correlation of GEBV vs observed. A multi-trait selection index assigned scores based on optimal ranges (heading date 80–90 days, plant height 115–125 cm, grain shape 3.0–3.4) and top percentiles (>80%, etc.) for yield components. All 58,353 pseudo-combinations among 367 restorer and 159 male sterile lines were scored. Validation combined 58 inbred lines with 19 commercial female lines to create 1,102 novel combinations; 67 were field-tested (Shanghai 2021), comparing top 20% vs bottom 20% by index. Web resource and code: A web portal (http://ricehybridresource.cemps.ac.cn/#/) provides access to samples, phenotypes, variants, and GWAS. Code is available on GitHub and Zenodo.

Phenotypic trends across breeding periods: Compared to Y1, Y3 hybrids showed a ~3-day reduction and reduced variance in heading date; increased valid panicle number, full grain number per panicle and per plant; longer and wider flag leaves; and markedly improved grain quality (cooking and appearance), including decreased chalkiness and increased gel consistency. Grain shape elongated over time. Genomic diversity and japonica introgression: Nucleotide diversity increased over breeding periods. WA (three-line) and TJ (two-line) hybrids were genetically differentiated; TJ hybrids increased in proportion and carried higher levels of japonica introgression than WA. Genome-wide maps revealed heterogeneous patterns of homozygous and heterozygous japonica-origin segments. Key genes with japonica-favorable alleles introgressed included NAL1 (leaf size, spikelet number), Waxy (grain quality), GW3p6 and GS6 (grain size/weight). TJ hybrids more frequently carried favorable alleles at these loci than WA, indicating new germplasms broadened resources. Breeding signatures via GWAS: In indica-indica hybrids, accumulation of favorable alleles increased from Y1 to Y3 for leaf size, grain number, yield, and quality traits. For heading date, extreme counts of alleles for short heading date decreased in Y3, consistent with tighter agronomic targeting. Specific loci included: LL1p6 (leaf length; PVE 3.75%; d/a ≈ 0.15, near-additive), LL9q21 (leaf length; partial dominance, d/a ≈ 0.40), NAL1 proximal signal (leaf width PVE 12.71%; grain number per panicle PVE 8.11%; d/a ≈ 0.61 and 0.79), GNP3p1 (grain number per panicle; PVE 1.89%; negative dominance, d/a ≈ −0.14), and GT8p4 (grain translucency PVE 14.09%; chalkiness PVE 6.76%; d/a ≈ 0.35 and 1.24). Major quality genes (Waxy, ALK, GW5/GSE5, GS3, GW7/GL7) showed increasing favorable allele frequencies over time. Most quality-related loci with PVE >1% exhibited negative dominance effects, aligning with the observation that quality traits seldom display heterosis; breeders therefore fixed favorable homozygotes by improving both parents. A yield–quality tradeoff at GW5/GSE5 was managed by discarding the higher grain weight allele that deleteriously affected quality. Intersubspecific heterosis: Among 68 indica-japonica F1 hybrids, 65.47% of genomic segments were indica/japonica heterozygous. In F2 mapping (5,342 individuals), 64 significant signals were detected; 55 resided in regions where indica/japonica genotypes were present in >50% of hybrids. Well-known divergently selected genes (DEP1, Ghd7, Ghd8, Hd1) were implicated, indicating widespread genetic complementation. In Quanjingyou No.1, four major loci (near sd1/bin0357, Hd1/bin1687 from maternal; bin0796; DEP1/bin2547 from paternal) mapped within complementary regions with d/a values 1.40, 0.24, 0.09, 0.01; 76% of F2s with heterozygotes at all four loci exceeded both parents in grain number per plant, evidencing transgressive segregation driven by complementation. Across eight indica-japonica F2 populations, QTLs with partial dominance and overdominance were balanced in count, but partial dominance QTLs contributed larger PVE for most yield traits (except seed setting rate); for kilo-grain weight, positive partial dominance QTLs predominated. Genomic selection: A GBLUP model trained on 12,678 individuals achieved prediction accuracies: grain shape 0.945; heading date 0.799; plant height 0.790; seed setting rate 0.728; full grain number per panicle 0.689; valid panicle number 0.559; yield per plant 0.518. An A+D model did not improve accuracy over additive-only. A selection index integrating seven traits scored 58,353 pseudo-combinations, identifying high-diversity, high-scoring combinations not yet bred. In a validation set of 67 novel combinations, top 20% by index outperformed bottom 20% for yield per plant (P = 2.17×10^-3) and full grain number per panicle (P = 7.54×10^-4), and showed reduced variance in seed setting rate. High-scoring WA combinations accumulated more favorable alleles across top-100 GWAS peaks and prior QTNs for yield traits and avoided sterility genotypes at rf3, rf4 (WA) and tms5 (TJ).

This comprehensive genomic and phenotypic dissection of rice hybrids across five decades reveals that modern hybrid breeding simultaneously broadened genetic diversity and improved performance by pyramiding favorable alleles and refining parental lines. Unlike the genetic bottlenecks observed in inbred cultivar improvement, hybrid breeding diversified germplasm via distinct mating systems (three-line vs two-line), notably increasing japonica introgression in two-line (TJ) hybrids, which provided access to favorable alleles at key loci such as NAL1, Waxy, and grain size genes. Phenotypically, breeders achieved a delicate balance: slightly shortening heading date for stability under variable climates while enlarging source (flag leaf) and sink organs (panicles) and enhancing grain quality, thereby maintaining or modestly improving yield while meeting quality standards. Genetic analyses clarify that for indica-indica hybrids, many key loci are near-additive or exhibit partial dominance; for quality traits, negative dominance predominates, necessitating fixation of favorable homozygotes through parental improvement. In indica-japonica hybrids, widespread complementation at subspecies-differentiated loci underlies strong heterosis; while both partial dominance and overdominance contribute, partial dominance loci tend to explain more variance for major yield components. These insights support breeding strategies that exploit heterozygous complementation where beneficial while eliminating deleterious dominant or negatively dominant effects. The genomic selection framework demonstrates robust, widely applicable prediction across diverse genetic backgrounds, aided by inclusion of large F2 populations capturing transgressive segregations and offsetting limited indica-japonica F1 representation. Although non-additive effects are pivotal for heterosis, including dominance in GBLUP did not improve prediction, echoing broader findings that reliable estimation of heterogeneous dominance effects remains challenging. Nonetheless, the multi-trait selection index effectively prioritizes combinations with high breeding potential and filters against incompatible sterility genotypes, offering a practical tool for accelerating hybrid design.

By resequencing and phenotyping 2,839 rice hybrids and 9,839 F2s, this study maps breeding footprints, quantifies heterotic loci and their dominance properties, elucidates the genetic basis of intersubspecific heterosis via genome-wide complementation, and delivers a validated genomic selection model and selection index for designing optimal hybrid combinations. Key contributions include: (1) demonstrating increased genomic diversity in hybrids over time driven by new male-sterility systems and japonica introgression; (2) identifying major loci and their effect modes (additive, partial dominance, negative dominance) underlying improvements in yield components and grain quality; (3) showing that partial dominance loci contribute more to variance than overdominance in indica-japonica yield heterosis; and (4) providing a high-accuracy, broadly applicable genomic prediction and selection pipeline that correlates with favorable allele accumulation and avoids sterility incompatibilities. Future work should expand intersubspecific hybrid datasets, incorporate multi-environment phenotyping, refine models to capture non-additive and epistatic effects, and integrate causal gene knowledge to enable molecular design of heterosis.

The F1 indica-japonica hybrid subset was relatively small compared to indica-indica hybrids, potentially limiting power to detect intersubspecific effects in F1s (mitigated by large F2 cohorts). Phenotyping and validation were conducted in specific environments in China (Hangzhou 2020; Shanghai 2021), which may limit environmental generalizability. While non-additive effects are important in heterosis, adding dominance to the GBLUP model did not increase prediction accuracy, reflecting current limitations in reliably estimating heterogeneous dominance effects across loci; epistasis was not explicitly modeled. Field validation of the selection index involved 67 combinations, providing supportive but limited-scale empirical validation.