Joint analysis of phenotype-effect-generation identifies loci associated with grain quality traits in rice hybrids

Rice grain quality is increasingly important to consumers and encompasses milling, appearance, cooking/eating, and nutritional traits. While hybrid rice often has superior yield compared to inbred varieties, its grain quality frequently requires improvement. Genetic dissection of hybrid quality is complicated by nonadditive effects such as dominance, and effective mapping ideally integrates data from both parents and hybrids while modeling additive and dominant components jointly. This study addresses the challenge by proposing a joint analysis of phenotypes, effects, and generations (JPEG) to simultaneously model additive and dominant genetic effects using phenotypes from inbred parental lines and their test-cross hybrids, aiming to identify loci associated with key grain quality traits and enable prediction of superior hybrid crosses.

Extensive prior work has cloned and characterized genes affecting grain size and quality in inbred rice, such as GS3, GL3.1, GW7/GL7 for grain length; GW2, GW5/qSW5, GS5 for grain width; GS6, GS9, TGW6, and GW8/SPL16 influencing grain size. Chalkiness is influenced by genes like OsRab5a and Chalk5. Cooking and eating quality is largely driven by amylose content controlled by the Wx gene and its alleles (e.g., Wxa, Wxb), and gel temperature is associated with OsSSIIa. In breeding contexts, dominance and other nonadditive effects complicate hybrid trait dissection. Joint additive–dominant analyses in other species (e.g., cattle) have revealed dominant loci of large effect, and large-scale hybrid mapping in maize has partitioned additive, dominance, and epistatic QTL. However, a unified framework that integrates parental and hybrid phenotypes while modeling additive and dominant effects jointly has been lacking, motivating the JPEG approach used here.

Germplasm and field design: Following a diallel-like testcross design, 115 indica accessions were used as male parents and crossed to five male-sterile lines as female parents, generating 565 (113×5 after excluding two with missing phenotypes) hybrid test crosses. Field trials were conducted in Wuhan, China (2013) in randomized blocks with two replicates. Twelve grain quality traits were phenotyped for male parents and hybrids: grain length (GL), grain width (GW), GL/GW ratio (GLWR), chalkiness degree (CD), percentage of grains with chalkiness (PGWC), transparent degree (TD), amylose content (AC), alkali spreading value (ASV), protein content (PC), brown rice ratio (BRR), milled rice ratio (MRR), and head rice ratio (HRR). Three representative plants per plot were harvested and evaluated after standard post-harvest handling. Genotyping and SNP processing: Genomic DNA from 118 parental varieties was sequenced on Illumina HiSeq2500 (average ~11×). Reads were quality filtered; alignment to IRGSP-1.0 (Nipponbare) used BWA; PCR duplicates removed by SAMtools. SNPs were called with GATK UnifiedGenotyper (diploid), filtered for missing rate <20% and MAF ≥5%, yielding 1,619,588 SNPs. Missing genotypes were imputed with NPUTE. Hybrid genotypes were inferred from parental genotypes. Additive genotypes were coded 0/1/2 (two homozygotes/heterozygote), and dominant genotypes were coded 0/1 (homozygotes/heterozygote). Population structure and LD: Phylogeny (SNPhylo), PCA (GAPIT), and ADMIXTURE (optimal K=5) characterized structure in parents and hybrids. LD was assessed with PLINK (pairwise r2), showing slow decay (r2 >0.3 within ~200 kb), consistent with rice. Phenotype derivations: Four phenotype types were formed: V (inbred parental observations), T (hybrid test-cross observations), G (general combining ability per parent; mean of its test-crosses minus overall hybrid mean), and H (heterosis per cross; hybrid minus mid-parent). Each of V, T, G, and H was standardized before analysis. Joint GWAS (JPEG pipeline): For combined analysis, observations (V, T, G, H) were vertically concatenated into a single phenotype vector Y. Matching genotype matrices for additive (A) and dominant (D) components were constructed, with D set to all zeros for V and G. A and D matrices were horizontally concatenated. Multiple-locus GWAS was performed with BLINK (GAPIT v3), iteratively selecting SNPs with significant additive or dominant effects as covariates along with PCs and female-parent dummies. Bonferroni-corrected 1% type I error threshold was applied. Separate analyses were also run for inbred-only (V and G), hybrid-only (T and H), and combined datasets. Heritability and prediction: Additive and dominant kinships (VanRaden) were computed from A and D. Mixed models (BGLR, RKHS) estimated additive, dominant, and total (additive+dominant) heritabilities in hybrids; additive heritability was also estimated in inbreds. Predictive ability was assessed via twofold cross-validation: hybrids were split into halves; GWAS was conducted only on the training half plus all inbreds; identified additive and dominant SNPs were used as predictors in a gBLUP framework (kinship from A and D) to predict the testing half. This was repeated with roles swapped for 100 replicates. Classical gBLUP using all markers served as a baseline.

- Phenotypic distributions: Most traits were approximately normal in inbreds; ASV (left-skewed) and PGWC (right-skewed) deviated. CD and AC showed bimodality. Hybrids generally fell between parental values, consistent with additive effects, though female-parent rank varied by trait. - Trait correlations: Strong positive correlations between CD and PGWC across V, T, G, H (r≈0.84–0.93). GL and GW were moderately negatively correlated (r≈−0.42 to −0.57). GL had moderate negative correlations with CD and PGWC; GW had moderate positive correlations with them. PC and ASV were weakly correlated with other traits, suggesting potential for independent improvement. - SNP discovery and loci: GWAS identified 192 significant SNPs that clustered into 128 loci (merging within 200 kb). By data source, 14, 68, and 89 loci were identified in inbred-only, hybrid-only, and combined analyses, respectively. Genetic effect classes: 44 additive loci, 97 dominant loci, and 13 loci exhibiting both additive and dominant effects. Eight of the 13 dual-effect loci were detected in all three analyses. - Known genes: Many loci mapped within 200 kb of known genes affecting grain quality, including Wx (AC), GS3 (GL, GLWR), GW5 (GW, GLWR), BG1 (Big Grain1; GW and GL in hybrid/combined), and OsGRF8 (GLWR in hybrids; AC in combined). The top AC signal corresponded to Wx and matched the observed bimodal phenotype distributions. - Trait-wise mapping: Milling quality traits yielded 1 significant locus in inbreds, 19 in hybrids, and 24 in combined analyses. A total of 17, 20, and 18 loci were identified for GL, GW, and GLWR, respectively; GLWR shared loci with both GL (three) and GW (six), with some loci shared among all three. - Variance explained and prediction: The identified additive and dominant loci together explained more than 30% of the genetic variation for hybrid performance for each trait. In cross-validation, pyramiding detected loci achieved prediction accuracies corresponding to over 45% of total heritability for all traits except HRR (≈30%), and outperformed standard gBLUP, demonstrating utility for selecting superior hybrid crosses.

The study demonstrates that jointly modeling additive and dominant effects across both parental inbreds and F1 hybrids increases GWAS power and interpretability for hybrid rice grain quality. The JPEG pipeline leverages multiple-locus modeling (BLINK) to iteratively control confounding and capture both effect types, identifying numerous loci including known quality genes (e.g., Wx, GS3, GW5) and additional candidates. The strong CD–PGWC correlation confirms shared genetic control of chalkiness metrics, whereas weak correlations of PC and ASV suggest they can be improved with minimal trade-offs. Importantly, loci identified via JPEG enable meaningful genomic prediction of hybrid performance and facilitate the identification of superior crosses from existing parental genotypes and partial hybrid phenotypes. Despite the dominance focus, only a small subset of loci exhibited both additive and dominant effects on the same trait, and overdominance could not be clearly dissected, likely due to limited statistical power. The framework can in principle be extended to epistasis (AA, AD/DA, DD), but computational burden rises sharply. Overall, integrating generations and effect types advances the genetic dissection of hybrid quality and supports practical breeding decisions.

This work introduces the JPEG pipeline for joint analysis of phenotypes, effect classes (additive and dominant), and generations (inbreds and hybrids) in hybrid rice. Applying JPEG to 12 grain quality traits in 113 male inbreds, five female testers, and 565 hybrids identified 128 associated loci, including many near known quality genes. These loci collectively explain substantial fractions of genetic variance in hybrid performance and markedly improve genomic prediction relative to gBLUP, enabling the identification of superior test crosses. Future research should increase sample sizes and tester diversity to enhance power, specifically to resolve overdominance, and extend JPEG to incorporate epistatic interactions while maintaining computational feasibility. The approach and code (open-source) are broadly applicable to hybrid breeding programs beyond rice.

- The analysis was limited to additive and dominant effects; epistatic interactions were not modeled due to computational constraints, potentially missing important loci. - Overdominance could not be clearly resolved: only 13 loci showed both additive and dominant effects and none exhibited clear overdominance for a given trait, indicating insufficient power to fully dissect dominance architectures. - Trait TD in female parents was below average, influencing hybrid TD unfavorably; while informative, this may limit generalizability for that trait in this panel.