Vol.:(0123456789) Theoretical and Applied Genetics (2026) 139:52 https://doi.org/10.1007/s00122-026-05159-z ORIGINAL ARTICLE Enhancing genomic prediction ability of blast resistance using genome‑wide association study‑derived marker weights in two rice (Oryza sativa L.) populations Félicien Akohoue1   · Cristian Camilo Herrera1 · Silvio James Carabali Balanta1 · Juanita Torres1 · Constanza Quintero1 · Gloria Mosquera1 · Maria Fernanda Alvarez1 Received: 3 June 2025 / Accepted: 11 January 2026 © The Author(s) 2026 Abstract Key message  Leaf and panicle blast resistances were moderately correlated and controlled by several genes, includ- ing Pi2/Pi9 and Pi33. GWAS-based marker weighting increased GBLUP predictive ability by up to 37% across two rice populations. Abstract  Breeding for blast resistance remains a high priority in rice (Oryza sativa L.) improvement, yet the genetic com- plexity of leaf blast (BL) and panicle blast (PB) continues to challenge prediction accuracy in genomic selection (GS). Traditional GS approaches, such as genomic best linear unbiased prediction (GBLUP), assume equal contribution from all markers, potentially limiting the capture of key resistance loci. Recent advances integrating genome-wide association studies (GWAS) into GS offer new opportunities to weight markers based on their biological relevance. In this study, we dissected the genetic architecture of BL and PB resistance in two diverse rice populations and evaluated the performance of three weighted GBLUP models that incorporate marker information from GWAS. Marker weighting strategies included FST-based weighting (FST-w), squared additive effects (AE-w), and − log10(p)-based weighting (− log10(p)-w). We identified signifi- cant marker-trait associations (MTAs), including key loci near the Pi2/Pi9 cluster and Pi33 gene regions on chromosomes 6 and 8. A moderate genetic correlation (0.43–0.44) between BL and PB severity suggests partially shared genetic control. Across traits and populations, AE-w and − log10(p)-w models improved predictive ability by 4–37% (0.03–0.23) and reduced normalized root mean square error by 3.8−35.3% relative to the unweighted GBLUP. These results demonstrate the value of integrating GWAS into GS (GS + GWAS) and highlight marker weighting as a practical strategy to enhance prediction accuracy for complex traits like blast resistance, ultimately accelerating genetic gains in rice breeding programs. Introduction Rice (Oryza sativa L.) represents a major staple food crop for more than half of the world population and serves as a primary source of calories and livelihoods in many coun- tries, particularly across Asia, the Americas, and Africa (FAO 2023). Despite its vital contribution to global food security, rice production faces several biotic and abiotic stresses which have a significant impact on yield and grain quality across production regions (Shah et al. 2019; Liu et al. 2021; Radha et al. 2023). Major stresses include blast dis- eases caused by the plant-pathogenic fungus Magnaporthe oryzae B.C. Couch (syn. Pyricularia oryzae) which affects yield and grain quality negatively worldwide (Perez-Nadales et al. 2014; Liu et al. 2021). The pathogen infects above- ground tissues at different growth stages, causing lesions on leaves, panicles and other organs (Ghatak et al. 2013). Under field conditions, blast infections can result in an aver- age yield loss of 10 to 30%, with a total crop failure in severe conditions (Fisher et al. 2012; Perez-Nadales et al. 2014; Asibi et al. 2019; Devanna et al. 2022). The fungus can sur- vive on infected residues for several rice cultivation cycles, which represent a primary source of inoculum (Raveloson et al. 2018). Integrated approaches comprising appropriate agronomic practices and high host resistance represent the most effective management strategy for blast disease (Asibi Communicated by Joshua N. Cobb. * Félicien Akohoue F.Akohoue@cgiar.org 1 Rice Program, International Centre for Tropical Agriculture (CIAT), Alliance Bioversity and CIAT, Americas Hub, Km 17 Recta Cali–Palmira, CP 763537 Palmira, Colombia http://crossmark.crossref.org/dialog/?doi=10.1007/s00122-026-05159-z&domain=pdf http://orcid.org/0000-0002-2160-0182 Theoretical and Applied Genetics (2026) 139:52 52   Page 2 of 27 et al. 2019). Cultivating blast-resistant varieties is a prefer- able control measure to reduce the use of fungicides and their harmful environmental effects. Resistance to blast disease is generally classified into two categories, namely complete and quantitative resist- ance. Complete resistance, also called qualitative resistance, is controlled by a single race-specific major resistance (R) gene, which is effective against specific M. oryzae strains possessing the corresponding avirulence genes (Koizumi 2010). About 100 Pyricularia genes (Pi genes) have been reported, of which 25 genes have been cloned and widely used by diverse breeding programs (Li et al. 2019). These include well known Pi genes like Pi9, which have been used in many breeding programs (Rathour et al. 2016; Xiao et al. 2019; Zhou et al. 2020; Misman et al. 2021; Fengshun et al. 2024). However, the complete resistance is not durable, as it can easily break in the presence of a new pathogen race. On the other hand, quantitative resistance, also known as field resistance, is a non-race-specific and polygenic resist- ance which involves multiple small effect genes that allow pathogen infection but restrict lesion expansion, preventing the disease progression. Over 700–800 quantitative trait loci (QTLs) harboring blast resistance genes have been reported from diverse genetic backgrounds and environmental condi- tions (Tian et al. 2022; Devanna et al. 2024). Unlike com- plete resistance, partial resistance is more durable and does not break down within a few years (Srivastava et al. 2017), although its use by breeding programs through marker- assisted selection (MAS) is reduced due to their genomic complexity. The combination of the two types of resist- ance is ideal to develop rice varieties with long-lasting and broad-spectrum resistance for more effective control of blast disease in famers’ fields. Many worldwide grown cultivars with high broad-spectrum resistance have been reported, including Moroberekan, IR64 and Jao Hom Nin (Sallaud et al. 2003; Chaipanya et al. 2017). To effectively leverage these genomic loci for higher blast resistance, the integration of high-throughput genom- ics-assisted breeding methods into breeding programs is of paramount importance. Standard marker-assisted selection (MAS) has been successfully implemented to develop blast- resistant cultivars based on major-effect R genes (Jiang et al. 2019; Yang et al. 2019; Liu et al. 2024). However, the effec- tiveness of MAS in leveraging small-effect genes for a dura- ble resistance is limited due to the extensive marker develop- ment efforts and costs involved. With the ever-decreasing genotyping costs and the availability of dense genome-wide single nucleotide polymorphism (SNP) arrays, advanced molecular methods such as genome-wide association stud- ies (GWAS) and genomic selection (GS) have emerged as promising solutions to improve breeding progress (Ahmadi 2022; Bartholomé et al. 2022). GS offers the potential to use all available genomic loci through robust prediction models, thereby enhancing breeding efficiency and genetic gains. In rice, GS potential has been investigated for many traits, such as yield and yield components (Grenier et al. 2015; Spindel et al. 2016; Wang et al. 2017; Zhang et al. 2023) and other agronomic traits (Grenier et al. 2015; Onogi et al. 2015; Zhang et al. 2023). Bartholomé et al. (2024) suggested that genomic selection could be useful for predicting tolerance to aluminum toxicity in upland rice, with a prediction accuracy of 0.67 for days to flowering, 0.60 for plant height, 0.53 for yield, and 0.65 for zinc. Unfortunately, the application of GS to predict blast resistance remains limited. Recently, Huang et al. (2019) evaluated 323 African and USDA accessions under artificial infections with 10 blast strains and reported genomic prediction accuracies ranging from 0.15 to 0.72. With this variability of prediction accuracy across strains, further investigations are necessary to fully harness the potential of GS for durable blast resistance. Studies should focus on both types of blast disease and expand to broader genetic backgrounds and environments. Several genomic prediction methods have been proposed, with genomic best unbiased linear predictor (GBLUP) being among the most applied models (Jeon et al. 2023; Montesinos-López et al. 2024a). GBLUP directly estimates genomic estimated breeding values of genotypes without explicitly estimating individual marker effects. Unlike tradi- tional BLUP models that use pedigree information, GBLUP replaces the pedigree with a genomic relationship matrix (G-matrix), built based on genome-wide markers like sin- gle nucleotide polymorphisms (SNPs) (VanRaden 2008; Su et al. 2012). The G-matrix captures the genetic similarity between genotypes and allows the GBLUP model to pre- dict their breeding values (Yang et al. 2014). In doing this, GBLUP model naively assumes that all available markers contribute equally to the genetic variation of the trait (Tiezzi and Maltecca 2015). This assumption ignores differences between markers and dilute contributions from major loci. However, not all genome-wide markers have the same influ- ence on traits like disease resistance. In practice, some mark- ers may be located near or within QTLs with large effects, while others may have little to no effect on the trait. Failing to prioritize these large effects markers could lead to reduced GBLUP prediction accuracy (Nishio and Satoh 2015; Zhang et al. 2016). To address this limitation, weighted GBLUP (wGBLUP) models that account for unequal marker contributions have been proposed and demonstrated to outperform the tradi- tional GBLUP model (Li et al. 2015; Nishio and Satoh 2015; Karaman et al. 2018; Gualdrón Duarte et al. 2020; Ren et al. 2021). wGBLUP incorporates markers' weights which are defined based on various weighting methods ranging from fixation index (FST)-based weights (Chang et al. 2019) to more biologically relevant approaches such as the incorpo- ration of GWAS results (Dong et al. 2016; Ren et al. 2021). Theoretical and Applied Genetics (2026) 139:52 Page 3 of 27  52 Despite its outperformance over the standard GBLUP, the genomic prediction accuracies of wGBLUP models are trait- dependent, as reported by Ren et al. (2021). To date, the impact of wGBLUP on genomic prediction accuracy of the different types of blast resistance remains unknown. There is a critical need to evaluate different weighting strategies for predicting blast resistance across diverse genetic back- grounds to provide practical insights to optimize genomic selection in rice breeding for blast resistance. This study aims to address these gaps by (i) investigat- ing the genetic architecture of leaf blast (BL) and panicle blast (PB) resistance using single-trait genome-wide associa- tion study (GWAS) within two distinct populations, namely SSD Tropics and 3K rice populations, and (ii) evaluating the genomic prediction ability of single-trait (ST) and multi-trait (MT) GBLUP models for BL and PB severity, incorporating three G-matrix weighting methods such as FST-based method (FST-w), squared additive effects (AE-w) and negative base 10 logarithm of P-value (− log10(p)-w) across both popula- tions. The use of cross-population comparative approach in this study was pivotal for validating the robustness and con- sistency of the G-matrix weighting methods across various genetic backgrounds, ensuring their applicability in diverse rice breeding programs targeting blast resistance. Materials and methods Plant materials The study included two rice populations: single seed descent (SSD) Tropics and 3K populations. SSD Tropics popula- tion comprised 24 interconnected families developed from crosses between ten restorer lines (Table 1) selected based on their moderate to high blast resistance. F1 families were advanced to F6 using the rapid generation advance (RGA), generating 1,484 lines with 20 to 88 lines per family (Table 1). On the other hand, the 3K population was com- posed of 204 accessions randomly selected from the global 3K germplasm (Supplementary file 1). Selected accessions belonged to “ind1A” (seven), “ind1B” (three), “ind2” (four), “ind3” (91), “indx” (95), “trop1” (two), “aromatic” (one) and “admix” (one) variety types. Most accessions (199) were distributed across diverse Asian subregions (Supplementary file 1). Field evaluation Both populations were evaluated under field conditions with a natural source of blast inoculum at the Fedearroz Santa Rosa experimental station at Villavicencio, Colom- bia. SSD Tropics lines were evaluated in 2023 and 2024 with four checks, including Olimar, FEDEARROZ-2000, FED-ITAGUA and ORYZICA-1. Lines were evaluated using an augmented design with row-column adjustment in 2023, and an augmented design without row-column adjustment in 2024. In each experiment, each check was included with 61 replicates. The 3K population was evaluated with the same checks from 2020 to 2023 using an alpha lattice design with two replicates. The following data were collected: days to flowering (DF, days), leaf blast (BL) severity, panicle blast (PB) severity and plant vigor (PV). DF was recorded plot-wise when at least 50% of plants flowed. BL was collected plot-wise at 30 to 40 days after planting using the Standard Evalua- tion System of IRRI (2014). BL typically initiates as small lesions which begin near the leaf tips or margins and extend downward, turning from pale green to yellow. Lesions can sometimes cover the entire leaf in susceptible varieties, while severe infections cause wilting and plant death. Like BL, PB was recorded on the panicle at 21–25 days after flowering using the same scale. PB is characterized by dark, necrotic lesions that partially or completely cover the panicle base, upper internode, or lower panicle axis, resulting in grayish panicles with partially filled or unfilled grains. To ensure a homogenous pressure of blast pathogen across the entire experiment, the highly susceptible cultivar Fanny was mixed with other indica susceptible genotypes to be used as spreader rows planted across the field. Genotypes with Table 1   The 10 restorer lines used as parents for developing SSD Tropics population, and their leaf blast (BL) and panicle blast (PB) resistance status Each female parent was crossed with the four male parents, generat- ing 24 interconnected families Genotype Blast status BL PB a. Female parent: FEM1 Highly resistant Moderately resistant FEM2 Moderately resistant Moderately resistant FEM3 Moderately resistant Moderately resistant FEM4 Highly resistant Highly resistant FEM5 Highly resistant Highly resistant FEM6 Highly resistant Highly resistant b. Male parent: MAL1 Moderately resistant Highly resistant MAL2 Highly resistant Highly resistant MAL3 Highly resistant Highly resistant MAL4 Moderately resistant Moderately resistant c. Checks: Olimar Highly susceptible Highly susceptible FEDEARROZ-2000 Highly susceptible Moderately resistant FED-ITAGUA​ Highly resistant Highly resistant ORYZICA-1 Moderately resistant Moderately resistant Theoretical and Applied Genetics (2026) 139:52 52   Page 4 of 27 a disease score lower than or equal to 3 were considered as resistant. PV was collected plot-wise in the SSD Tropics population using a scale of 1–9 scale, with 1 being ‘‘excel- lent vigor’’ and 9 being ‘‘very poor vigor.’’. Genotyping and marker filtering All “SSD Tropics” lines and their parents were genotyped using the 1k-RiCA v4.2 single nucleotide polymorphism (SNP) array (Arbelaez et al. 2019). In total, 1,094 SNP markers were obtained, including 261 trait markers, 28 purity markers, and 805 genome-wide markers. The marker data were filtered by removing SNPs with a minor allele frequency (MAF) lower than 5% and missing values greater than 20%. The remaining missing data, representing 0.56% of the SNPs, were imputed using Wright’s equilibrium method (Wright 1922). After filtering and imputation, 671 high-quality SNPs were retained for downstream analyses. For the 204 lines from the 3K population, one million SNP markers were obtained from the public Rice SNP-Seek database (https://​3kric​egeno​me.​s3.​amazo​naws.​com/​3kRG_​ downl​oad.​html) (Mansueto et al. 2017). The same filtering criteria were applied, reducing the dataset to 431,377 SNP markers. The marker data were further refined and narrowed by implementing a selective linkage disequilibrium prun- ing (SLDP) adapted from the procedure described by Zhu et al. (2023). LD was calculated using the squared allele frequency correlation adjusted for kinship relationships ( r2 v  ) (Mangin et al. 2012): where r2 v (i, j) is the kinship-adjusted LD estimate between markers i and j; Xv is the kinship-adjusted genotype matrix. Xv was calculated as follows: where X is the genotype matrix and K is the kinship matrix calculated using the VanRaden method (VanRaden 2008). Briefly, SLDP was implemented in four steps, which involved: (1) identifying significant SNPs through a GWAS analysis for each trait and their highly linked ( r2 v  ≥ 0.95) neighbors within a 50 kb window; (2) pruning significant SNPs and linked neighbors based on r2 v  ≥ 0.95 and GWAS P values; (3) conducting genome-wide pruning of remaining SNPs using r2 v  ≥ 0.80; and (4) combining pruned SNPs to produce a final dataset containing all GWAS-detected SNPs. The SLDP process reduced the marker dataset to 9,126 high- quality SNP markers for the 3K population. (1)r2 v (i, j) = (Cov(Xv i ,Xv j ))2 Var ( Xv i ) .Var ( Xv j ) (2)Xv = K−1∕2X Phenotypic data analysis Phenotypic data were analyzed using a two-stage approach to account for differences in experimental design appropri- ately. From the three repeated BL scorings, only the high- est score was used for each genotype to assess performance under highest blast pressure while minimizing the poten- tial effect of weather variability which could influence dis- ease pressure during field evaluations. In the first stage, the analysis was performed separately per environment (loca- tion × year combinations), and adjusted means were esti- mated for each genotype and their respective Smith weight was estimated. Given the ordinal nature of the traits, except for DF, a cumulative logit mixed model was fitted for each trait within each population. For SSD Tropics, the mixed model was fitted as follows: where yiknt is the response of genotype i in row n and column t within block k; P(yiknt ≤ c) is the probability of yiknt being in category c or below; αc is the category specific threshold (intercept); gi is the genotype effect; bk is the block effect; wnk is the effect of row n within block k; ltk is the effect of column t within block k. In the 3K population, the model was fitted as follows: where yijk is the response of genotype i in block k within replicate j; gi is the genotype effect, rj is the replicate effect; bjk is the effect of block k within the replicate j. Models (1) and (2) were fitted following the Bayesian approach with four chains using the BRMS R package (Bürkner 2017). For each chain, total Markov Chain Monte Carlo (MCMC) iterations, warmup and thinning were set to 20,000, 6000, and 2, respectively. To control for the potential effect of flowering date on panicle blast severity, DF was included as a covariate in the first-stage PB model. Unlike ordinal scale data (BL, PB and VG), DF was analyzed using the Gaussian link family following Eq. (3) and (4). where εiknt and εijk are residual errors. Model evaluation was done based on Gelman-Rubin diagnostic statistics such as effective sample size (ESS) and R-hat ( R̂ ) (Gelman et al. 2004, 2014). ESS measures the number of independent samples after accounting for auto- correlation between draws. An ESS ≥ 400 indicates more reliable estimates and better convergence. For the combined (3)SSD ∶ logit [ P ( yiknt ≤ c )] = �c + gi + bk + wnk + ltk (4)3K ∶ logit [ P ( yijk ≤ c )] = �c + gi + rj + bjk (5)SSD ∶ yiknt = gi + bk + wnk + ltk + �iknt (6)3K ∶ yijk = gi + rj + bjk + �ijk https://3kricegenome.s3.amazonaws.com/3kRG_download.html https://3kricegenome.s3.amazonaws.com/3kRG_download.html Theoretical and Applied Genetics (2026) 139:52 Page 5 of 27  52 chains, ESS value was calculated for each model parameter (Vehtari et al. 2021) as follows: where N is the number of post-warmup draws, M is the number of chains, ρt is the autocorrelation at lag t, and k is the truncation point used to limit the sum of autocorre- lations to reduce noise. The best k was chosen following the initial positive sequence estimator method, where the sum was truncated at the largest k, for which all autocor- relations remained positive. N was estimated for each chain as follows: Moreover, R̂ assesses the convergence of the different Markov chains. It compares between- and within-chain variances to determine if the chains have mixed well and converged to the posterior distribution. R̂ was calculated following the improved procedure by Vehtari et al. (2021) as follows: At R̂ = 1 chains are perfectly converged and well-mixed, and the posterior distribution has been sufficiently explored. v̂ar + (�|y) is the marginal posterior variance which com- bines both between- (B) and within-chain (W) variances as follows: ESS and R̂ were rank-normalized to improve convergence diagnostics as recommended by Vehtari et al. (2021). In all models, genotype was fitted as fixed effects and corresponding adjusted means were estimated using the pos- terior_linpred() function. For each genotype and adjusted mean, a weight was estimated as the diagonal elements of the inverse of the variance–covariance (Vj) matrix following the Smith weighting method as described by Möhring and Piepho (2009): where SWj is the Smith weight of each genotype in environ- ment j and Vj is the variance–covariance matrix of genotypes in environment j. (7)ESS = NM 1 + 2 ∑2k+1 t=1 �t (8)N = MCMCiterations - warmup thinning (9)R̂ = √ �var + (𝜃|y) W (10)v̂ar + (�|y) = N − 1 N W + 1 N B (11)SWj = D(V−1 j ) In the second stage, the following mixed linear mod- els were fitted for each trait using the ASReml-R package (Butler et al. 2023): where yijk and yik are adjusted mean from first stage for each genotype; gi is the genotype effect; fj is the family effect; ek is the environment effect; geik is the effect of genotype- by-environment interaction; fejk is the effect of family-by- environment interaction; and εijk and εik are residual errors. Smith weight estimated from the first stage was incorporated into the second stage model to separate the genotype-by- environment interaction and residual variances. Genotype, family and environment were fitted as random effects to esti- mate variance components for each trait. The likelihood ratio test was performed to evaluate the statistical significance of variance components. In addition, genotype was fitted as fixed effects to estimate best linear unbiased estimates (BLUE) for each genotype across environments. Broad sense heritabil- ity (H2) was estimated as follows (Piepho and Möhring 2007): where �2 g  is the genotypic variance, vΔ is the variance of a difference between two BLUEs. To estimate genotypic correlation between traits in each population, the second stage models were extended to a bivariate model described as follows: where y1, and y2 are adjusted means of genotype for the first and second trait, respectively. Bivariate models were fitted with a heterogeneous variance–covariance structure using corgh option for genotype, family, environment and residual. Based on BLUEs from the second stage, the phe- notypic diversity within each population was further described by performing a hierarchical cluster analy- sis using the FactoMineR package (Lê et al. 2008). All analyses were performed in the R software 4.4.2 (R Core Team 2024). (12)SSD ∶ yijk = gi + fj + ek + geik + fejk + �ijk (13)3K ∶ yik = gi + ek + geik + �ik (14)H2 = �2 G �2 G + vΔ 2 (15)SSD ∶ [ y1 y2 ] = gi + fj + ek + geik + fejk + �ijk (16)3K ∶ [ y1 y2 ] = gi + ek + geik + �ik Theoretical and Applied Genetics (2026) 139:52 52   Page 6 of 27 Population structure and marker‑trait association analysis Based on the high-quality markers (671 SNPs for SSD Trop- ics and 9,126 SNPs for the 3K), population structure was investigated using principal components analysis. Individual admixture coefficients were estimated for all lines, assuming 1 to 13 ancestral populations (K), to determine the probabil- ity of each genotype to be included in a distinct subpopula- tion. For each K, 40 iterations were performed to estimate cross-entropy values that were useful to determine the opti- mal number of subpopulations. A genotype was included in a specific subpopulation when its inclusion probability was greater than 60%. Genotypes that did not meet this inclusion criterion were considered admixed. Admixture analysis was done using the LEA package (Frichot and François 2015). To evaluate the influence of family structure on the genetic differentiation within SSD Tropics population, an analysis of molecular variance (AMOVA) was performed using the poppr.amova() function from the poppr R package (Kamvar et al. 2014). The analysis applied the pegas method (Paradis 2010) to partition genetic variation across three hierarchical levels: between populations (defined by family), between genotypes within family, and residual (i.e., heterozygosity). Within each population, the genetic architecture of BL and PB was investigated by conducting a genome-wide asso- ciation analysis. A mixed linear model (MLM) that incor- porated kinship and population structure (Wang and Zhang 2021) was fitted as follows: where y is the vector of BLUE for each genotype; β is the fixed effect including grand mean and population structure; α is the marker effect; u ~ N(0, 2 K �2 a  ) is a vector of size n (number of individuals) for random polygenic effects; ε ~ N(0, I �2 �  ) is a vector of random residual effects. X, W and Z are design matrices for β, α and u, respectively. K is the kinship matrix calculated using the VanRaden method (VanRaden 2008). MLM was fitted using the Genomic Association and Prediction Integrated Tool (GAPIT) pack- age v.3.1.0 (Wang and Zhang 2021). Significant marker-trait associations (MTAs) were identified based on a corrected Bonferroni threshold of 10−5 and 10−6 in SSD Tropics and 3K populations, respectively. The threshold was determined in each population by dividing the α value of 0.05 by the number of markers. LD block analysis was performed per chromosome based on all markers to identify quantitative trait loci (QTL) regions which were associated with the traits. LD block was identified using a modified version of the block partition method referred to as Big-LD by Kim et al. (2018), by replacing the standard r2 estimate by the (17)y = X� +W� + Zu + � kinship-adjusted squared correlation ( r2 v  ) to compute LD as previously described in Eq. (1), followed by a graph- based clique detection using a threshold of r2 v  ≥ 0.80. The Genome Annotation Project database (RGAP; https://​ rice.​uga.​edu) (Hamilton et al. 2025) was queried against physical positions of GWAS-detected markers on the Nipponbare reference genome Os-Nipponbare-Refer- ence-IRGSP1.0 to retrieve gene ontology and descrip- tion for most significant marker-trait associations for both traits. Unweighted genomic relationship matrix construction To implement single-trait and multi-trait genomic predic- tion for BL and PB, additive genomic relationship matrices (G-matrix) were constructed using the AGHmatrix v2.1.4 package (Amadeu et al. 2023) following the method proposed by VanRaden (2008). The unweighted G-matrix (Gunw) was calculated as: where Z is a n × m matrix (n = number of genotypes, m = number of SNPs) which contains SNP genotype coeffi- cients at each SNP. The coefficients of SNP i with alleles A1 and A2 are 0–2pi for homozygous allele A1 (A1A1), 1 − 2pi for the heterozygous state (A1A2), and 2–2pi for homozy- gous allele A2 (A2A2), where qi and pi are the frequencies of A1 and A2, respectively. Weighted genomic relationship matrices construction Weighted additive G-matrices (Gw) were calculated for each prediction model as follows: where w is a diagonal matrix with the ith diagonal element being SNP weight at locus i. SNP weights were determined using three methods: the FST-based weighting method (FST-w), squared additive marker effect-based method (AE-w) and negative base 10 logarithm of GWAS P value-based method (− log10(p)-w). In FST-w, the genotypic matrix was weighted by weights derived from FST values of individual SNPs. FST values were calculated based on population structure results, using the SNPRelate package following Weir and Cockerham (1984): (18)Gunw = ZZ� 2 ∑ piqi (19)Gw = ZWZ� 2 ∑ piqi (20)FST = a a + b + c https://rice.uga.edu https://rice.uga.edu Theoretical and Applied Genetics (2026) 139:52 Page 7 of 27  52 where a is the variance of allele frequencies among sub- populations; b is the covariance of allele frequencies within subpopulations and c is the average expected heterozygosity within subpopulations. For SNP j, the relative weight was calculated as described by Chang et al. (2019): where wj is the weight for SNP j, FSTj is the FST value of SNP j and n is the number of SNPs. Moreover, AE-w and − log10(p)-w are GWAS statistics-derived weighting methods. Weights were calculated and scaled for both meth- ods following the procedure described by Su et al. (2014) as in Eq. (22) and (23) as follows: Marker effect-derived weight: GWAS P value-derived weight: where AEj is additive effect of SNP j derived from GWAS analysis, pj is the frequency of the alternative allele of SNP j. To fit multi-trait prediction models, SNP weights were calculated as a weighted linear combination of trait-specific marker weights (wj). The combined weight, denoted as cwj, was calculated using the following formula: where α1 and α2 are trait-specific scaling coefficients that account for genetic correlation between trait1 and trait2, wj,trait1 is weight of SNP j for trait1 and wj,trait2 is weight of SNP j for trait2. α1 and α2 are defined based on the genetic correlation between the two traits, such that α1 + α2 = 1. We calculated α1 and α2 using Eqs. (25) and (26) as follows: where rg is the genetic correlation between the two traits. Here, as the genetic correlation increases, the contributions of the individual traits to the combined weight become more balanced, maximizing the weights of shared loci. This enables the second trait—typically the one with lower (21)wj = FSTj∑n j=1 FSTj n (22)wj = 2pj(1 − pj)AE 2 j ∑n j=1 wj n (23)wj = −2pj(1 − pj) log10(P − value) ∑n j=1 wj n (24)cwj = �1wj,trait1 + �2wj,trait2 (25)�1 = 1 1 + rg (26)�2 = rg 1 + rg heritability—to leverage the effects of shared loci more effectively when the correlation is high. Genomic prediction analysis and models evaluation Single-trait (ST) and multi-trait (MT) GBLUP models were fitted in the ASReml-R package (Butler et al. 2023) for BL and PB severity based on the BLUE from the phenotypic analysis and each of the genomic relationship matrices. The ST model was fitted for each trait as follows: Y is the vector of BLUE values corresponding to the genotypes; 1 is the vector with elements 1; µ is the grand mean effect; ZA is design matrix that associates the genomic breeding values with the response variable; gA is the vector of genomic breeding values; and ε is the residual term. gA ~ N(0, G σ2 g ); µ ~ N(0, I σ2 ε ); cov(ε, gA) = 0. G is the additive genomic relationship matrix. MT model was fitted by extending the ST model to a bivariate model as follows: where [ Y1 Y2 ] is the vector of BLUEs for trait1 and trait2; I1 and I2 are the identity matrices; [ �1 �2 ] is the vector of grand mean effects for trait1 and trait2; [ gA1 gA2 ] is the vector of the genomic breeding values of the two traits; ZA1 and ZA2 are design matrices that associate genomic breeding values with the response variables; [ �1 �2 ] is the vector of residual effects of the two traits . [ gA1 gA2 ]  ~ N(0, G  ⊗ H) , where H = [ �2 g1 �g12 �g12 �2 g2 ] is the variance–covariance matrix of the genomic breeding values of the two traits. [ �1 �2 ]  ~ N(0, I ⊗ R), where R = [ �2 �1 ��12 ��12 �2 �2 ] is the residual variance–covariance matrix. MT model was fitted with a heterogeneous vari- ance–covariance structure using corgh option for genotype and residual. The evaluation of Single-trait (ST) and Multi-trait (MT) models was performed using a fivefold cross-validation approach. This consisted in dividing the initial population into five folds, with each fold containing 296–302 genotypes for the SSD Tropics population and 37–45 genotypes for the 3K population. When composing the folds, a stratified sampling was applied to ensure that genotypes were selected from all (27)Y = 1� + ZAgA + � (28) [ Y1 Y2 ] = [ I1 0 0 I2 ][ �1 �2 ] + [ ZA1 0 0 ZA2 ][ gA1 gA2 ] + [ �1 �2 ] Theoretical and Applied Genetics (2026) 139:52 52   Page 8 of 27 families in the SSD Tropics population, and all subpopula- tions in the 3K population. The stratified sampling accounts for population structure and optimizes the representativeness of the training set. Additionally, each fold was sampled 100 times, and a leave-one-fold-out approach was used to compose the training and validation sets. At each iteration, one fold was excluded and used as the validation set, while the remaining four folds were combined to form the training set. This process was repeated until each of the five folds served as the valida- tion set. This ensures that every genotype was included in the validation set exactly once, while the remaining data was used for model training. This allowed for a comprehensive evalua- tion of the model performance across the entire population to minimize overfitting and ensure independency of the results from a specific data partition. At each iteration and for each validation set, the model predictive ability was determined by calculating the Pear- son correlation between genomic estimated breeding values (GEBV) and BLUE of each genotype in the validation set. To further evaluate the performance of genomic prediction models, the root mean square error (RMSE) was calculated for each validation set and iteration. RMSE was normalized by the standard deviation of each trait, yielding the normal- ized RMSE (nRMSE) to enable meaningful comparison across weighting methods and traits. For each genotype i, nRMSE was defined as the square root of the squared difference between the genomic estimated breeding value (GEBVi) and the observed phenotype (yi), divided by the standard deviation (SD) of the trait: where N is the number of genotypes in each validation set. (29)nRMSE = � 1 N ∑N i=1 (GEBVi − yi) 2 SD Prior to constructing G-matrices, GWAS analyses were conducted for each training set at every iteration to calcu- late marker weights for the AE-w and − log10(p)-w weight- ing methods. This approach prevents information leakage between the training and validation sets and minimizes biases in model evaluation, especially for the weighted models. Results Genetic variation within each population for all traits Model diagnostic statistics showed perfect convergence (maximum Ȓ = 1), with good chains mixing (Table 2). Fur- thermore, all traits had a high minimum effective sample size (ESSmin > 400). This also suggests high precision in parameter estimates, reducing the uncertainty associated with marginal means. In both populations, the genetic variance was substan- tial for all traits (Table 2). Genotype-by-environment inter- action variance also contributed to the trait variability, although it was smaller than the genetic variance. Her- itability estimates were high for all traits, ranging from 0.63 to 0.82, suggesting that a significant portion of the phenotypic variance was attributable to genetic factors. In SSD Tropics, within-family variance ( σ2 G  ) was 1.2 − two- fold higher than between-family variance for all traits. The density plot within each population showed that BLUE values for all traits were continuously and relatively nor- mally distributed (Fig. 1). Table 2   Stage 1 model diagnostic statistics, and variance components and heritability estimates from stage 2 DF days to 50%  flowering (days), BL  severity = leaf blast severity, PB  severity = panicle blast severity, VG = plant vigor, σ2 G  = genotypic variance, σ2 F  = family variance, σ2 GE  = genotype-by-environment interac- tion variance, σ2 FE  = family-by-environment interaction variance, σ2 ε  = residual variance, R̂max = R-hat con- vergence diagnostic, ESSmin = minimum effective sample size and ESSmax = maximum effective sample size Gelmn-Rubin statistic Variance components and heritability Trait R̂max ESSmin ESSmax σ2 G σ2 F σ2 GE σ2 FE σ2 ε H2 a. SSD Tropics: BL severity 1.00 727 20,084 0.32 0.26 0.12 0.09 0.01 0.81 DF 1.00 822 6235 21.35 10.81 0.01 0.01 2.90 0.78 PB severity 1.00 955 2294 0.29 0.25 0.24 0.03 0.02 0.66 VG 1.00 785 6480 0.71 0.52 0.25 0.11 0.10 0.63 b. 3K population: BL severity 1.00 10,232 12,387 2.41 0.06 0.80 0.83 DF 1.00 6540 20,349 152.79 4.42 0.48 0.82 PB severity 1.00 8578 16,044 0.40 0.01 0.55 0.69 Theoretical and Applied Genetics (2026) 139:52 Page 9 of 27  52 Genetic correlation and clustering in SSD Tropics and 3K populations In both SSD Tropics and 3K populations, positive and mod- erate genetic correlations (0.43–0.44) were detected between BL and PB severity (Figs. 2a and 3a). Correlations between disease severity (BL and PB) and agronomic traits such as DF and VG were generally low. The multivariate analysis revealed that the first two components explained 66.0% and 80.5% of the phenotypic diversity in SSD Tropics and 3K populations, respectively. BL and PB severity were corre- lated with Dimension 1, while DF and VG were correlated with Dimension 2 (Figs. 2b and 3b). The hierarchical clustering identified three distinct clus- ters (Figs. 2c and 3c) in the two populations. In the SSD Tropics population, Cluster 1 showed the lowest average BL severity (2.52), while Cluster 2 had the lowest average PB severity (2.19) (Fig. 2d). On the other hand, Cluster 3 exhib- ited the highest average severity for BL (4.50) and PB (3.56). Similar results were obtained in the 3K population (Fig. 3d). Cluster 1 showed the lowest BL and PB severity, while the highest disease severity was observed with Cluster 3. Population structure revealed by principal component analysis The results of the population structure analysis for the SSD Tropics population are illustrated in Fig. 4. Average cross- entropy values decreased steadily with K and plateaued around K = 10 (Fig. 4a). Based on the inclusion probability of 60%, the number of distinct subpopulations was 10, with several admixed genotyped (Fig. 4b). The PCA biplot shows that the first two principal components explain 10.1% and 7.7% of the total genetic variance, respectively (Fig. 4c). The clustering pattern was mainly explained by family, with most of them being overlapped. The AMOVA results indi- cated that a significant proportion of the total genetic varia- tion (42.5%) was attributed to differences between families (Table 3). Interestingly, a higher proportion of variation was observed among lines within families, representing 50.8% of the total genetic variation. In contrast, the residual variation SSD Tropics 3K 2.5 5.0 7.5 2.5 5.0 7.5 0.00 0.05 0.10 0.15 0.20 0.0 0.1 0.2 0.3 BL severity D en si ty SSD Tropics 3K 2 4 6 8 2 4 6 8 0.0 0.2 0.4 0.6 0.0 0.1 0.2 0.3 0.4 0.5 D en si ty SSD Tropics 3K 70 80 90 100 70 80 90 100 0.00 0.02 0.04 0.06 0.000 0.025 0.050 0.075 DF D en si ty SSD Tropics 2.5 5.0 7.5 0.0 0.1 0.2 0.3 VG D en si ty a b c d PB severity Fig. 1   Density plots showing the distribution of best linear unbiased estimates (BLUEs) for all traits within SSD Tropics and 3K popula- tions. a = leaf blast (BL) severity, b = panicle blast (PB) severity, c = days to 50% flowering (DF), and d = plant vigor (VG). VG was recorded in SSD Tropics population only. The blue dashed line indi- cates the population mean for each trait Theoretical and Applied Genetics (2026) 139:52 52   Page 10 of 27 within lines—primarily reflecting individual heterozygo- sity—was minimal (Table 3). In the 3K population, the cross-entropy curve showed an optimum of four subpopulations (K = 4) as shown in Fig. 5a. This was confirmed by the heatmap of admixture coefficients (Fig. 5b) based on the inclusion probability of 60%, which showed four distinct subpopulations and many admixed genotypes. Additionally, the PCA revealed that the first two principal components clearly discriminated the four identified subpopulations, and explained 6.7% and 3.6% of genetic variance, respectively (Fig. 5c). Marker‑trait associations for leaf blast and panicle blast severity The genetic architecture of BL and PB severity was depicted by implementing a single-trait genome-wide association study within each population. Ten and four subpopulations were included in the GWAS model as covariates to account for population structure in SSD Tropics and 3K populations, respectively. Significant marker-trait associations were iden- tified on chromosomes 6 and 8 in the SSD Tropics popula- tion and chromosomes 1, 6 and 12 in the 3K population (Fig. 6, Supplementary file 2). In SSD Tropics population, several significant MTAs were identified on chromosomes 6 and 8 for BL severity and chromosome 6 for PB severity. LD block analysis revealed several linkage blocks with r2 v  ≥ 0.8 delimiting QTL regions, including qtl6.6 and qtl8.3 on chromosomes 6 and 8, respec- tively. Within qtl6.6, common MTAs were detected for BL and PB severity, of which the most significant MTA was MSU7_6_10388389_TT-AA on chromosome 6 (Table 4). This common MTA explained about 2% of the phenotypic variation. MTAs detected on chromosome 8 within qtl8.3 explained about 19% of phenotypic variance each. In the 3K population, MTAs were detected on chromosomes 6 and 12 for BL severity and chromosome 1 for PB severity (Table 4). QTL regions, namely qtl1.6, qtl6.18 and qtl12.17 were defined by significant markers on chromosomes 1, 6 and 12, respectively. The most significant MTA for BL severity −0.10 0.10 −0.01 0.17 0.43** −0.11 DF VG VG BL DF VG BL −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 Dim 1 (39.7%) D im 2 (2 6. 3% ) Cor −1.0 −0.5 0.0 0.5 1.0 −3 −1 1 3 −3 −1 1 3 D im 2 (2 6. 3% ) Cluster 1 2 3 *** *** 0 2 4 6 8 10 BL D is ea se s ev er ity Cluster 1 2 3 Dim 1 (39.7%) a b c d PB PB PB PB Fig. 2   Description of phenotypic diversity within SSD Tropics popu- lation based on days to 50% flowering (DF), leaf blast (BL) severity, panicle blast (PB) severity and plant vigor (VG). a = genetic corre- lation between traits, b = correlation between traits and the first two dimensions, c = clustering of the genotypes on the first two dimen- sions, d = statistical difference between clusters for BL and PB sever- ity. *** significant at p < 0.001 Theoretical and Applied Genetics (2026) 139:52 Page 11 of 27  52 in this population was 191472769 within qtl6.18, detected on chromosome 6 with 6% of explained phenotypic variance. Gene ontology search in the Rice Annotation Project Database (RAP-DB) identified several candidate genes within QTL regions (Supplementary file 3). On chro- mosome 6, common MTAs (MSU7_6_10388389_TT-A, MSU7_6_10389352_A-T, Pi2-01, Pi2-02) for BL and PB severity from SSD Tropics population and the most signifi- cant MTA (191472769) for BL severity from 3K population were all located within locus Os06g0286700 (Table 5). This locus represents Pi2/Pi9 gene cluster which encodes a nucle- otide-binding site leucine-rich repeat (NBS-LRR) protein. In addition, MTA Pi33_3 specifically detected for BL in SSD Tropics population was linked to Pi33 gene which encodes avirulence conferring enzyme 1 (ACE1)-specific protein. Based on the most significant SNP markers linked to Pi2/ Pi9 gene cluster and Pi33_3, haplotype groups with low average disease severity (< 3) were identified within each population (Fig. 7). With Pi2/Pi9, average BL and PB sever- ity in best haplotypes was reduced by 1.1−2.6 and 0.4−0.8 across the two populations, respectively, compared to groups with highest disease severity (Fig. 7a, b). Similarly, average BL severity of best haplotype group (BL severity ≈ 2.68) for Pi33 was 0.97 lower than the group with highest severity (BL severity ≈ 3.65) (Fig. 7c). Predictive ability of weighted and unweighted genomic prediction models In the SSD Tropics population, average predictive ability ranged from 0.78 to 0.81 for BL severity (Fig. 8a) and 0.62 to 0.67 for PB severity (Fig. 8b). The highest predictive abil- ity was observed for both traits with − log10(p)-w and AE-w, while FST-w and the unweighted models exhibited the lowest values. Average predictive abilities of BL severity with both AE-w and − log10(p)-w were 0.03 higher compared to the unweighted models, and 0.03−0.05 higher for PB severity (Fig. 8a, b). No difference in the average predictive ability was observed between FST-w and the unweighted models. Similarly, − log10(p)-w and AE-w showed no significant −0.08 0.44* −0.14 BL BL DF DF BL −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 Dim 1 (52.3%) D im 2 ( 28 .2 % ) −1.0 −0.5 0.0 0.5 1.0 −3 −1 1 −3 −1 1 3 D im 2 ( 28 .2 % ) Cluster 1 2 3 *** *** 0 2 4 6 8 10 BL D is ea se s ev er it y Cluster 1 2 3 Dim 1 (52.3%) a b c d PB PB PB Fig. 3   Description of phenotypic diversity within 3K population based on  days to 50%  flowering (DF), leaf blast (BL) severity and panicle blast (PB) severity. a = genetic correlation between traits, b = correlation between traits and the first two dimensions, c = clus- tering of the genotypes on the first two dimensions, d = statistical dif- ference  between clusters for BL and PB severity. *** significant at p < 0.001 Theoretical and Applied Genetics (2026) 139:52 52   Page 12 of 27 difference within this population for both traits. Average predictive abilities of ST and MT models were statistically similar for all methods. The average predictive ability in the 3K population ranged from 0.74 to 0.90 for BL severity (Fig. 8c) and 0.62 to 0.85 for PB severity (Fig. 8d). Like in the SSD Tropics population, the highest average predictive abilities in the 3K population were observed with − log10(p)-w and AE-w for both traits (Fig. 8c, d). The unweighted models consistently exhibited the lowest predictive abilities (Fig. 8c). Depend- ing on ST and MT models, the average predictive ability for BL severity with AE-w was 0.14–0.16 higher com- pared to the unweighted models, and 0.14–0.23 higher for PB severity. Similarly, − log10(p)-w showed an increase of 0.12–0.14 in the average predictive ability for BL severity and 0.12–0.20 for PB severity (Fig. 8c, d). In contrast, the increase in predictive ability from FST-w was smaller, with 0.03 for both traits. A significant difference was observed between ST and MT models for AE-w and − log10(p)-w for both traits in the 3K population. For BL severity, the aver- age predictive ability of ST model was 0.02 higher than that of MT model with AE-w and − log10(p)-w. For PB severity, ST model outperformed MT model by 0.09 and 0.08 with AE-w and − log10(p)-w, respectively. In contrast, the predic- tive abilities of ST and MT models were similar for FST-w and unweighted models. Based on the unweighted models, the predictive ability of BL severity in the SSD Tropics population was 0.04 higher compared to the 3K population, while similar values were observed for PB severity between the two populations. 0.40 0.45 0.50 0.55 0.60 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Number of populations (K) C ro ss −e nt ro py a −8 −6 −4 −2 0 2 4 6 8 −8 −6 −4 −2 0 2 4 6 8 10 PC1 (10.1%) PC 2 (7 .7 % ) CT26676 CT26677 CT26678 CT26679 CT26680 CT26681 CT26682 CT26683 CT26684 CT26685 CT26686 CT26687 CT26688 CT26689 CT26690 CT26691 CT26692 CT26693 CT26694 CT26695 CT26696 CT26697 CT26698 CT26699 Parent Tester c 0.00 0.25 0.50 0.75 1.00 G en 12 05 G en 12 19 G en 12 37 G en 12 61 G en 12 95 G en 12 66 G en 11 97 G en 12 21 G en 12 88 G en 12 89 G en 12 86 G en 12 56 G en 12 96 G en 12 45 G en 28 0 G en 12 72 G en 12 17 G en 12 35 G en 12 24 G en 12 04 G en 12 83 G en 12 48 G en 13 17 G en 12 97 G en 11 31 G en 12 02 G en 12 85 G en 12 49 G en 12 74 G en 12 93 G en 12 03 G en 12 54 G en 13 18 G en 12 08 G en 12 42 G en 13 06 G en 12 31 G en 12 40 G en 12 34 G en 12 25 G en 13 14 G en 12 16 G en 12 60 G en 12 65 G en 12 58 G en 11 69 G en 11 95 G en 12 68 G en 12 52 G en 12 53 G en 12 11 G en 11 80 G en 12 64 G en 11 93 G en 12 70 G en 11 88 G en 12 80 G en 12 55 Fe m 5 G en 13 38 G en 11 74 G en 99 6 G en 11 34 G en 11 07 G en 96 2 G en 11 15 G en 10 65 G en 10 39 G en 11 17 G en 14 52 G en 12 79 G en 98 2 G en 13 91 G en 11 19 G en 11 11 G en 11 42 G en 12 99 G en 11 25 G en 11 46 G en 13 73 G en 13 98 G en 94 4 G en 13 72 G en 13 54 G en 11 24 G en 13 10 G en 12 67 G en 93 0 G en 11 02 G en 12 81 G en 10 98 G en 95 0 G en 12 51 G en 13 09 G en 14 27 G en 13 68 G en 11 70 G en 10 99 G en 12 98 G en 11 13 G en 10 12 G en 12 63 G en 13 21 G en 13 90 G en 13 92 G en 95 9 G en 21 50 G en 12 75 G en 10 71 G en 11 72 G en 11 26 G en 97 7 G en 12 76 G en 11 48 G en 13 56 G en 10 67 G en 11 75 G en 13 77 G en 94 5 G en 97 6 G en 11 33 G en 13 61 G en 93 6 G en 13 04 G en 13 07 G en 94 6 G en 10 41 G en 11 79 G en 10 94 G en 12 43 G en 14 21 G en 12 92 G en 12 27 G en 11 41 G en 10 31 G en 11 28 G en 97 9 G en 13 30 G en 14 28 G en 11 01 G en 13 35 G en 12 50 G en 14 20 G en 13 89 G en 13 50 G en 11 45 G en 11 16 G en 12 10 G en 93 1 G en 11 43 G en 11 27 G en 13 31 G en 11 37 G en 11 32 G en 13 12 G en 94 2 G en 13 36 G en 12 78 G en 93 7 G en 13 45 G en 14 24 G en 98 5 G en 11 35 G en 13 55 G en 11 10 G en 10 38 G en 11 68 G en 10 68 G en 21 81 G en 13 76 G en 12 15 G en 97 8 G en 97 0 G en 11 83 G en 99 2 G en 11 61 G en 14 29 G en 93 5 G en 64 8 M al 1 G en 98 1 G en 19 42 G en 19 07 G en 19 00 G en 15 26 G en 18 74 G en 01 6 G en 36 5 G en 10 75 G en 10 64 G en 31 1 G en 18 80 G en 00 9 G en 93 9 G en 66 7 G en 29 9 G en 18 68 G en 19 52 G en 93 4 G en 01 1 G en 68 2 G en 15 10 G en 97 4 G en 09 0 G en 17 95 G en 37 9 G en 30 1 G en 26 7 G en 36 6 G en 37 7 G en 35 5 G en 96 8 G en 31 5 G en 66 4 G en 28 5 G en 01 9 G en 33 8 G en 19 11 G en 18 71 G en 38 5 G en 19 03 G en 04 7 G en 15 65 G en 15 01 G en 29 3 G en 66 0 G en 28 4 G en 10 93 G en 92 9 G en 31 0 G en 03 7 G en 31 3 G en 18 03 G en 15 45 G en 30 6 G en 09 7 G en 65 1 G en 01 5 G en 16 59 G en 09 1 G en 18 10 G en 19 25 G en 18 93 G en 96 6 G en 36 7 G en 93 2 G en 34 6 G en 34 4 G en 15 63 G en 04 0 G en 93 3 G en 18 70 G en 18 84 G en 15 73 G en 34 3 G en 18 94 G en 15 85 G en 93 8 G en 34 7 G en 18 83 G en 19 45 G en 16 43 G en 33 3 G en 18 15 G en 01 4 G en 37 1 G en 67 9 G en 19 51 G en 19 26 G en 15 66 G en 01 3 G en 18 73 G en 19 01 G en 05 3 G en 10 73 G en 18 79 G en 16 44 G en 97 3 G en 36 4 G en 69 4 G en 30 4 G en 70 9 G en 15 00 G en 97 5 G en 30 3 G en 14 74 G en 18 53 G en 38 4 G en 99 3 G en 15 90 G en 33 0 G en 03 6 G en 64 0 G en 19 46 G en 17 89 G en 98 0 G en 18 54 G en 04 4 G en 05 4 G en 19 16 G en 10 48 G en 09 2 G en 18 85 G en 15 61 G en 15 46 G en 95 4 G en 99 4 G en 18 78 G en 67 1 G en 19 08 G en 96 4 G en 10 66 G en 29 7 G en 65 2 G en 14 64 G en 10 30 G en 14 98 G en 33 1 G en 15 68 G en 18 75 G en 94 3 G en 10 37 G en 10 69 G en 18 57 G en 35 3 G en 07 9 G en 00 7 G en 28 2 G en 15 49 G en 37 8 G en 97 1 G en 37 3 G en 18 89 G en 15 27 G en 19 05 G en 37 5 G en 36 3 G en 98 3 G en 97 2 G en 15 02 G en 92 8 G en 29 5 G en 34 5 G en 31 6 G en 96 5 G en 96 3 G en 18 20 G en 37 4 G en 32 9 G en 94 0 G en 14 61 G en 15 04 G en 26 2 G en 25 7 G en 76 6 Fe m 6 G en 84 8 G en 77 1 G en 72 1 G en 73 4 G en 92 7 G en 77 9 G en 77 3 G en 80 9 G en 78 7 G en 90 2 G en 87 4 G en 89 0 G en 76 1 G en 74 2 G en 81 8 G en 68 7 G en 78 4 G en 85 2 G en 64 4 G en 78 3 G en 81 9 G en 63 8 G en 73 3 G en 69 5 G en 68 3 G en 72 4 G en 75 2 G en 88 1 G en 81 5 G en 74 3 G en 77 4 G en 67 3 G en 70 6 G en 86 2 G en 88 0 G en 89 2 G en 76 0 G en 70 5 G en 86 7 G en 64 6 G en 74 6 G en 74 8 G en 80 0 G en 78 0 G en 83 3 G en 69 6 G en 67 0 G en 71 1 G en 71 2 G en 71 3 G en 79 7 G en 70 2 G en 77 8 G en 65 3 G en 85 1 G en 84 0 G en 75 1 G en 88 3 G en 67 7 G en 68 4 G en 64 7 G en 83 0 G en 69 1 G en 76 8 G en 74 9 G en 65 9 G en 91 6 G en 81 7 G en 65 6 G en 79 4 G en 66 3 G en 91 4 G en 80 3 G en 82 5 G en 73 8 G en 79 0 G en 81 6 G en 70 4 G en 71 8 G en 87 3 G en 83 8 G en 75 3 G en 83 7 G en 64 2 G en 69 3 G en 70 8 G en 73 0 G en 75 7 G en 75 4 G en 73 1 G en 64 3 G en 79 5 G en 83 6 G en 77 7 G en 88 8 G en 83 2 G en 79 9 G en 78 6 G en 82 7 G en 84 5 G en 81 4 G en 65 8 G en 90 5 G en 90 7 G en 82 4 G en 13 13 G en 91 9 G en 70 1 G en 69 7 G en 81 2 G en 72 2 G en 80 8 G en 89 7 G en 68 9 G en 92 4 G en 67 2 G en 77 5 G en 77 6 G en 82 6 G en 91 7 G en 89 5 G en 67 5 G en 65 5 G en 76 2 G en 87 0 G en 84 1 G en 66 1 G en 71 0 G en 75 0 G en 84 6 G en 92 6 G en 76 9 G en 87 9 G en 64 9 G en 89 6 G en 70 3 G en 70 7 G en 85 6 G en 80 7 G en 72 8 G en 90 3 G en 65 4 G en 69 8 G en 86 0 G en 75 9 G en 73 5 G en 66 9 G en 75 8 G en 84 9 G en 91 1 G en 89 8 G en 88 5 G en 64 1 G en 86 1 G en 84 7 G en 87 8 G en 85 0 G en 64 5 G en 72 3 G en 89 1 G en 80 5 G en 92 5 G en 77 2 G en 74 1 G en 92 3 G en 63 9 G en 83 5 G en 91 0 G en 72 6 G en 90 0 G en 91 8 G en 88 6 G en 13 80 G en 76 4 G en 67 4 G en 70 0 G en 87 5 G en 17 51 IR 80 55 9A .1 G en 16 97 Fe m 4 G en 17 60 G en 16 24 G en 16 17 G en 15 30 G en 17 75 G en 14 84 G en 15 51 G en 15 57 G en 16 45 G en 16 87 G en 18 56 G en 16 92 G en 15 52 G en 59 9 G en 15 62 G en 16 50 G en 16 86 G en 19 48 G en 16 80 G en 17 85 G en 15 06 G en 17 50 G en 17 52 G en 17 19 G en 16 89 G en 14 66 G en 14 68 G en 16 98 G en 84 4 G en 17 91 G en 16 10 G en 16 51 G en 15 67 G en 18 60 G en 17 92 G en 18 05 G en 17 00 G en 18 41 G en 18 46 G en 18 62 G en 15 05 G en 15 48 G en 16 79 G en 16 74 G en 15 70 G en 17 69 G en 14 67 G en 18 31 G en 18 61 G en 16 42 G en 19 50 G en 14 70 G en 17 29 G en 14 65 G en 17 84 G en 14 71 G en 15 87 G en 14 72 G en 16 46 G en 15 99 G en 14 58 G en 17 74 G en 17 27 G en 15 12 G en 15 11 G en 17 49 G en 16 33 G en 14 86 G en 17 99 G en 17 58 G en 16 55 G en 18 12 G en 16 73 G en 15 64 G en 16 88 G en 16 02 G en 17 70 G en 18 00 G en 15 38 G en 18 32 G en 16 19 G en 18 17 G en 15 14 G en 18 43 G en 17 65 G en 14 73 G en 17 97 G en 17 77 G en 16 08 G en 16 05 G en 15 28 G en 17 28 G en 14 83 G en 15 50 G en 16 70 G en 16 36 G en 16 26 G en 16 14 G en 17 83 G en 15 22 G en 16 90 G en 18 49 G en 16 76 G en 18 11 G en 16 96 G en 17 90 G en 15 59 G en 17 05 G en 16 28 G en 15 17 G en 81 1 G en 16 54 G en 17 59 G en 16 52 G en 16 64 G en 18 55 G en 15 07 G en 18 36 G en 15 18 G en 16 75 G en 14 62 G en 15 13 G en 14 69 G en 15 47 G en 17 09 G en 16 81 G en 16 69 G en 16 62 G en 16 53 G en 17 53 G en 17 22 G en 16 61 G en 16 47 G en 18 37 G en 17 11 G en 15 21 G en 16 95 G en 16 09 G en 17 62 G en 14 63 G en 15 94 G en 16 06 G en 18 59 G en 18 18 G en 18 50 G en 18 35 G en 17 66 G en 16 67 G en 16 11 G en 15 09 G en 16 99 G en 14 85 G en 18 19 G en 15 16 G en 18 25 G en 15 15 G en 17 64 G en 15 71 G en 16 21 G en 18 01 G en 24 8 G en 42 1 G en 40 1 G en 19 69 G en 44 5 G en 19 85 G en 19 70 G en 39 5 G en 20 26 G en 16 38 G en 16 18 G en 46 3 G en 03 2 G en 16 32 G en 11 0 G en 11 76 G en 15 92 G en 40 2 G en 19 27 G en 16 49 G en 11 08 G en 40 7 G en 16 57 G en 46 2 G en 20 36 G en 11 23 G en 48 0 G en 19 61 G en 16 15 G en 20 55 G en 16 16 G en 11 9 G en 41 8 G en 43 5 G en 16 07 G en 20 03 G en 03 5 G en 19 57 G en 41 9 G en 78 1 G en 16 20 G en 16 60 G en 15 98 G en 67 8 G en 47 5 G en 16 01 G en 44 7 G en 10 2 G en 11 18 G en 19 95 G en 43 3 G en 20 64 G en 19 99 G en 75 5 G en 47 4 G en 41 7 G en 08 3 G en 16 41 G en 43 1 G en 39 4 G en 19 65 G en 16 65 G en 47 6 G en 10 5 G en 11 00 G en 15 88 G en 73 7 G en 15 93 G en 39 6 G en 19 80 G en 11 50 G en 47 1 G en 20 62 G en 19 55 G en 42 6 G en 20 22 G en 67 6 G en 11 47 G en 65 7 G en 39 9 G en 41 1 G en 16 56 G en 55 0 G en 40 5 G en 77 0 G en 03 0 G en 19 93 G en 18 82 G en 16 63 G en 72 0 G en 16 13 G en 48 2 G en 16 29 G en 11 36 G en 11 86 G en 16 66 G en 19 71 G en 11 12 G en 16 40 G en 11 03 G en 20 37 G en 20 17 G en 19 79 G en 11 55 G en 76 3 G en 11 09 G en 73 9 G en 46 0 G en 73 6 G en 11 44 G en 74 7 G en 16 27 G en 11 81 G en 10 96 G en 20 32 G en 47 8 G en 11 4 G en 11 29 G en 48 3 G en 16 31 G en 11 82 G en 16 39 G en 16 68 G en 71 9 G en 16 23 G en 08 6 G en 72 7 G en 10 95 G en 14 57 G en 39 3 G en 44 3 G en 11 3 G en 45 7 G en 41 6 G en 40 6 G en 11 30 G en 40 9 G en 11 39 G en 11 49 G en 41 0 G en 19 66 G en 43 0 G en 20 01 M al 4 G en 11 87 G en 10 91 G en 19 49 G en 22 2 G en 23 7 G en 23 0 G en 27 9 G en 27 7 G en 24 6 G en 03 4 G en 20 6 G en 19 1 G en 05 1 G en 18 5 G en 14 3 G en 16 2 G en 19 0 G en 22 1 G en 04 9 G en 14 6 G en 21 1 G en 25 9 G en 12 1 G en 09 4 G en 15 1 G en 23 9 G en 24 3 G en 26 6 G en 18 1 G en 19 2 G en 14 0 G en 25 5 G en 11 2 G en 24 1 G en 19 8 G en 15 0 G en 25 2 G en 23 8 G en 20 2 G en 20 5 G en 20 0 G en 13 2 G en 22 4 G en 23 4 G en 02 6 G en 11 1 G en 02 4 G en 04 6 G en 11 7 G en 27 1 G en 25 3 G en 03 3 G en 03 8 G en 19 9 G en 00 8 G en 10 4 G en 10 3 G en 00 6 G en 00 2 G en 08 4 G en 08 5 G en 27 8 G en 12 0 G en 02 2 G en 25 8 G en 18 0 G en 21 9 G en 10 8 G en 18 2 G en 26 5 G en 04 2 G en 12 4 G en 02 3 G en 22 0 G en 13 6 G en 02 0 G en 02 5 G en 11 5 G en 01 2 G en 00 5 G en 16 6 G en 05 2 G en 18 7 G en 17 7 G en 04 3 G en 13 9 G en 17 4 G en 13 8 G en 10 1 G en 02 9 G en 00 3 G en 09 5 G en 01 7 G en 04 1 G en 13 7 G en 19 4 G en 17 5 G en 17 9 G en 12 3 G en 01 0 G en 19 5 G en 02 1 G en 09 8 G en 24 5 G en 02 7 G en 17 6 G en 09 3 G en 09 6 G en 17 2 Fe m 1 G en 03 9 G en 25 1 G en 20 1 G en 11 6 G en 24 9 G en 00 1 G en 04 5 G en 03 1 G en 04 8 G en 01 8 G en 23 2 G en 18 8 G en 25 0 G en 21 55 G en 21 90 G en 21 59 G en 21 54 G en 21 70 G en 21 64 G en 21 87 G en 21 49 G en 21 88 G en 21 80 G en 21 78 G en 21 71 G en 21 72 G en 21 77 G en 21 51 G en 21 63 G en 21 57 G en 21 37 G en 19 44 G en 21 44 G en 20 30 G en 21 46 G en 19 28 G en 19 92 G en 20 42 G en 19 54 G en 18 92 G en 18 04 G en 19 10 G en 18 14 G en 20 29 G en 19 04 G en 18 90 G en 20 27 G en 21 04 G en 21 06 G en 19 56 G en 19 68 G en 21 82 G en 21 75 G en 20 74 G en 19 38 G en 19 98 G en 20 45 G en 20 67 G en 18 13 G en 21 08 G en 20 25 G en 18 33 G en 20 66 G en 18 67 G en 19 97 G en 21 67 G en 20 07 G en 19 35 G en 21 76 G en 18 72 G en 21 68 G en 19 91 Fe m 3 G en 21 60 G en 21 52 G en 21 28 G en 20 34 G en 18 09 G en 18 96 G en 20 54 G en 21 79 G en 18 91 G en 20 11 G en 19 88 G en 19 47 G en 18 24 G en 19 78 G en 18 65 G en 19 90 G en 20 56 G en 21 45 G en 20 71 G en 21 03 G en 18 88 G en 18 69 G en 19 76 G en 21 21 G en 20 88 G en 20 00 G en 20 91 G en 19 31 G en 20 14 G en 21 74 G en 18 23 G en 21 94 G en 20 97 G en 20 77 G en 20 68 G en 21 31 G en 19 83 G en 21 27 G en 21 53 G en 19 77 G en 20 20 G en 20 16 G en 21 23 G en 18 98 G en 21 43 G en 18 95 G en 21 01 G en 20 10 G en 21 93 G en 21 95 G en 21 30 G en 19 53 G en 18 87 G en 19 96 G en 21 66 G en 20 35 G en 20 98 G en 18 81 G en 21 62 G en 20 15 G en 18 02 G en 21 85 G en 21 26 G en 21 02 G en 18 64 G en 20 02 G en 20 99 G en 19 67 G en 19 02 G en 20 85 G en 18 40 G en 18 58 G en 20 63 G en 20 76 G en 21 89 G en 20 04 G en 18 48 G en 21 58 G en 18 86 G en 19 75 G en 18 22 G en 21 48 G en 20 89 G en 18 76 G en 20 12 G en 21 47 G en 20 23 G en 18 07 G en 18 30 G en 19 86 G en 11 8 G en 18 21 G en 19 34 G en 18 45 G en 17 93 G en 20 93 G en 19 32 G en 19 06 G en 20 19 G en 20 33 G en 18 44 G en 14 10 G en 49 7 G en 18 26 G en 48 7 G en 50 8 G en 52 9 Fe m 2 G en 52 7 G en 52 6 G en 41 3 G en 42 4 G en 45 9 G en 51 5 G en 52 8 G en 48 1 G en 55 4 G en 51 9 G en 53 7 G en 50 5 G en 41 5 G en 30 5 G en 54 1 G en 44 9 G en 49 3 G en 63 1 G en 42 2 G en 35 2 G en 38 6 G en 56 8 G en 49 9 G en 53 2 G en 38 2 G en 53 0 G en 48 5 G en 58 9 G en 54 7 G en 49 0 G en 61 9 G en 53 3 G en 56 7 G en 42 8 G en 47 9 G en 44 1 G en 41 4 G en 48 9 G en 62 8 G en 57 2 G en 56 5 G en 35 6 G en 58 0 G en 42 0 G en 49 1 G en 57 1 G en 49 2 G en 42 9 G en 61 3 G en 43 8 G en 54 6 G en 58 8 G en 62 4 G en 56 6 G en 39 8 G en 38 7 G en 58 1 G en 42 5 G en 31 2 G en 38 3 G en 62 5 G en 52 1 G en 22 6 G en 55 1 G en 42 7 G en 40 8 G en 42 3 G en 55 7 G en 29 8 G en 51 0 G en 38 0 G en 43 6 G en 38 1 G en 37 2 G en 30 0 G en 37 0 G en 26 1 G en 57 9 G en 47 2 G en 24 4 G en 46 8 G en 35 7 G en 46 9 G en 63 6 G en 30 9 G en 46 6 G en 62 9 G en 36 9 G en 30 8 G en 36 8 G en 30 2 G en 44 8 G en 37 6 G en 31 4 G en 58 7 G en 33 2 G en 61 5 G en 49 6 G en 30 7 G en 60 5 G en 41 2 G en 45 5 G en 59 6 G en 76 7 M al 3 G en 13 20 G en 51 4 G en 20 4 G en 53 5 G en 51 3 G en 82 9 G en 81 3 G en 17 12 G en 50 9 G en 52 4 G en 18 9 G en 53 6 G en 12 44 G en 17 0 G en 49 8 G en 49 5 G en 21 2 G en 21 05 G en 11 38 G en 83 9 G en 81 0 G en 20 75 G en 48 8 G en 15 9 G en 13 0 G en 79 1 G en 17 1 G en 14 4 G en 20 79 G en 53 1 G en 16 1 G en 17 63 G en 20 8 G en 21 00 G en 12 18 G en 17 76 G en 17 55 G en 79 6 G en 16 77 G en 17 61 G en 20 7 G en 20 82 G en 50 6 G en 21 3 G en 21 19 G en 12 77 G en 16 94 G en 16 93 G en 12 73 G en 16 5 G en 20 73 G en 20 92 G en 16 4 G en 20 83 G en 50 7 G en 12 14 G en 17 71 G en 17 79 G en 16 78 G en 17 81 G en 50 1 G en 21 39 G en 20 70 G en 14 2 G en 17 48 G en 52 0 G en 22 3 G en 84 3 G en 12 9 G en 21 24 G en 21 15 G en 79 8 G en 20 84 G en 50 2 G en 82 0 G en 52 3 G en 50 4 G en 20 80 G en 80 4 G en 21 38 G en 11 94 G en 15 5 G en 82 1 G en 12 46 G en 17 73 G en 82 3 G en 48 4 G en 53 4 G en 52 5 G en 21 07 G en 21 32 G en 20 69 G en 17 14 G en 12 94 G en 17 57 G en 21 20 G en 82 2 G en 17 68 G en 17 02 G en 80 2 G en 18 6 G en 83 4 G en 51 8 G en 51 1 G en 80 6 G en 17 21 G en 80 1 G en 20 78 G en 16 9 G en 12 30 G en 21 41 G en 21 14 G en 21 25 G en 83 1 G en 84 2 G en 17 56 G en 13 4 G en 20 9 G en 21 09 G en 15 8 G en 23 5 G en 11 40 G en 20 72 G en 15 6 G en 21 29 G en 52 2 G en 12 71 G en 21 18 G en 21 16 G en 21 0 G en 21 13 G en 13 03 G en 13 1 G en 12 47 G en 17 8 M al 2 G en 88 9 G en 14 50 G en 89 3 G en 13 71 G en 59 0 G en 21 86 G en 13 79 G en 13 83 G en 59 8 G en 13 43 G en 57 3 G en 27 2 G en 13 32 G en 21 91 G en 17 94 G en 13 65 G en 26 4 G en 26 8 G en 91 3 G en 62 2 G en 22 7 G en 14 48 G en 18 27 G en 13 78 G en 26 3 G en 61 2 G en 61 6 G en 27 3 G en 55 3 G en 13 96 G en 56 0 G en 57 8 G en 21 84 G en 60 8 G en 21 56 G en 25 4 G en 90 8 G en 14 17 G en 13 39 G en 87 7 G en 60 6 G en 13 29 G en 86 5 G en 10 84 G en 24 2 G en 90 6 G en 86 8 G en 21 83 G en 62 1 G en 28 1 G en 21 69 G en 60 9 G en 22 8 G en 92 1 G en 13 42 G en 18 51 G en 13 62 G en 60 2 G en 14 26 G en 22 9 G en 13 84 G en 21 65 G en 86 6 G en 85 9 G en 13 85 G en 61 0 G en 18 08 G en 18 52 G en 18 28 G en 53 8 G en 61 8 G en 87 1 G en 26 0 G en 14 49 G en 86 4 G en 60 3 G en 63 5 G en 14 37 G en 19 12 G en 13 59 G en 88 7 G en 13 25 G en 13 53 G en 13 86 G en 85 3 G en 17 96 G en 23 3 G en 14 07 G en 60 7 G en 59 4 G en 22 5 G en 14 25 G en 13 24 G en 60 1 G en 54 5 G en 13 82 G en 14 47 G en 62 3 G en 24 7 G en 13 81 G en 63 2 G en 13 23 G en 13 37 G en 61 4 G en 91 2 G en 14 30 G en 62 7 G en 56 2 G en 13 41 G en 17 98 G en 63 4 G en 13 64 G en 55 2 G en 59 2 G en 59 5 G en 13 66 G en 55 5 G en 59 3 G en 27 5 G en 27 4 G en 21 92 G en 27 6 Genotype A nc es tr y co ef fic ie nt b Fig. 4   Principal component analysis of SSD Tropics lines, including the 1,484 F6 lines, their 10 parental lines based on 671 high quality SNP markers. a = Scree plot showing the optimum number of ancestral popula- tions; b = Heatmap of individual admixture coefficients for the optimum number of ancestral populations; c = Scatter plot illustrating genotypes groupings projected on the first two components Table 3   Results of analysis of molecular variance based on 671 SNP markers and the 1484 genotypes from the 24 families in the SSD Tropics population df degree of freedom Source of variation df Variance component % of total variance Between families 23 104.03 42.58 Among lines within family 1460 124.08 50.80 Within lines (residual) 1484 16.18 6.62 Total 2967 244.29 100.00 Theoretical and Applied Genetics (2026) 139:52 Page 13 of 27  52 Root mean square error of weighted and unweighted genomic prediction models In both populations, normalized root mean square error (nRMSE) was remarkably lower with all ST and MT weighted models than their corresponding unweighted ones (Fig. 9). For both traits, the lowest nRMSE values were observed with − log10(p)-w and AE-w, while the unweighted models showed the highest values. In SSD Tropics popula- tion, − log10(p)-w and AE-w reduced nRMSE by 0.06–0.07 (11.3%) for BL severity, and 0.03–0.05 (3.8–5.1%) for PB severity across both ST and MT models (Fig. 9a, b), while FST-w showed little to no nRMSE reduction. In addition, nRMSE reduction was higher in 3K popula- tion for all traits than the SSD Tropics population. In the 3K population, − log10(p)-w exhibited an nRMSE reduction of 0.17–0.20 (25–29.4%) for BL severity and 0.07–0.21 (7.5–26.9%) for PB severity across ST and MT models. (Fig. 9c, d). Similarly, AE-w reduced nRMSE by 0.21–0.24 (30.4–35.3%) for BL severity and 0.10–0.25 (9.7–32.1%) for PB severity. FST-w showed an nRMSE reduction of 0.03–0.17 (4.4–24.6%) for BL severity and 0.03–0.13 (3.8–14.0%). In both populations, nRMSE reduction by ST models was higher than that of MT models across traits and weighting methods, with the exception of FST-w in 3K population. Discussion With its assumption of equal contributions of all genetic markers to the trait of interest, the traditional GBLUP model has significant limitations that can lower genomic prediction accuracy (Nishio and Satoh 2015). By comparing three marker weighting methods such as FST-w, − log10(p)-w and AE-w using SSD Tropics and 3K populations, this study aimed to refine the genomic rela- tionship matrix and enhance predictive abilities for leaf blast and panicle blast resistance in rice. Compared to the standard unweighted approach, − log10(p)-w and AE-w exhibited the highest predictive ability with the lowest root mean square error for ST and MT models (Figs. 8 and 9). 0.37 0.38 0.39 0.40 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Number of populations (K) C ro ss −e nt ro py a −1.5 −1.0 −0.5 0.0 0.5 1.0 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 2.0 PC1 (6.7%) PC 2 (3 .6 % ) Admixed Pop1 Pop2 Pop3 Pop4 c 0.00 0.25 0.50 0.75 1.00 IR IS _3 13 −1 21 38 IR IS _3 13 −1 22 88 IR IS _3 13 −1 22 92 IR IS _3 13 −1 22 61 IR IS _3 13 −9 01 9 IR IS _3 13 −8 75 1 IR IS _3 13 −1 19 20 IR IS _3 13 −1 15 46 IR IS _3 13 −1 16 35 IR IS _3 13 −1 10 71 IR IS _3 13 −1 11 45 IR IS _3 13 −1 07 62 IR IS _3 13 −1 11 35 IR IS _3 13 −1 21 87 IR IS _3 13 −1 22 69 IR IS _3 13 −1 11 43 IR IS _3 13 −1 22 90 IR IS _3 13 −8 38 6 IR IS _3 13 −1 14 63 IR IS _3 13 −9 02 0 IR IS _3 13 −1 11 41 IR IS _3 13 −1 21 21 IR IS _3 13 −1 22 21 IR IS _3 13 −8 38 3 IR IS _3 13 −1 09 90 IR IS _3 13 −1 17 20 IR IS _3 13 −8 90 9 IR IS _3 13 −8 72 3 IR IS _3 13 −7 68 5 IR IS _3 13 −1 06 52 IR IS _3 13 −1 05 54 IR IS _3 13 −9 53 3 IR IS _3 13 −1 21 48 IR IS _3 13 −1 07 69 IR IS _3 13 −1 11 20 IR IS _3 13 −9 20 9 IR IS _3 13 −1 06 74 IR IS _3 13 −1 08 44 IR IS _3 13 −1 21 88 IR IS _3 13 −1 17 00 IR IS _3 13 −1 20 50 IR IS _3 13 −1 00 54 IR IS _3 13 −1 21 94 IR IS _3 13 −1 11 37 IR IS _3 13 −8 89 5 C X4 03 IR IS _3 13 −1 06 75 IR IS _3 13 −9 60 9 IR IS _3 13 −1 20 52 IR IS _3 13 −1 11 38 IR IS _3 13 −9 73 0 IR IS _3 13 −8 98 6 IR IS _3 13 −1 13 77 IR IS _3 13 −1 12 34 IR IS _3 13 −1 07 51 IR IS _3 13 −1 18 35 IR IS _3 13 −1 18 94 IR IS _3 13 −1 09 71 IR IS _3 13 −1 17 45 IR IS _3 13 −1 13 84 IR IS _3 13 −1 13 83 IR IS _3 13 −9 60 0 IR IS _3 13 −1 11 32 IR IS _3 13 −1 11 28 IR IS _3 13 −1 16 77 IR IS _3 13 −1 16 78 IR IS _3 13 −1 17 19 IR IS _3 13 −1 19 61 IR IS _3 13 −1 22 59 IR IS _3 13 −1 22 49 IR IS _3 13 −1 18 20 IR IS _3 13 −1 17 05 IR IS _3 13 −1 21 30 IR IS _3 13 −1 21 93 IR IS _3 13 −9 00 6 IR IS _3 13 −1 17 17 IR IS _3 13 −1 11 33 IR IS _3 13 −1 06 95 IR IS _3 13 −1 10 76 IR IS _3 13 −1 16 76 IR IS _3 13 −1 22 22 IR IS _3 13 −1 22 46 IR IS _3 13 −1 19 91 IR IS _3 13 −1 18 37 IR IS _3 13 −1 16 79 IR IS _3 13 −1 16 82 IR IS _3 13 −1 23 34 IR IS _3 13 −1 07 72 IR IS _3 13 −1 09 97 IR IS _3 13 −1 19 92 IR IS _3 13 −1 13 27 IR IS _3 13 −1 19 02 IR IS _3 13 −1 16 64 IR IS _3 13 −1 10 86 IR IS _3 13 −1 22 63 IR IS _3 13 −1 17 08 IR IS _3 13 −1 07 79 IR IS _3 13 −1 22 87 IR IS _3 13 −1 07 77 IR IS _3 13 −1 15 43 IR IS _3 13 −8 67 4 IR IS _3 13 −1 22 86 IR IS _3 13 −1 10 91 IR IS _3 13 −9 11 2 IR IS _3 13 −1 18 49 IR IS _3 13 −9 28 1 IR IS _3 13 −8 79 1 IR IS _3 13 −8 67 9 IR IS _3 13 −1 21 42 IR IS _3 13 −8 99 6 IR IS _3 13 −1 11 51 IR IS _3 13 −9 18 2 IR IS _3 13 −1 21 28 IR IS _3 13 −1 06 55 IR IS _3 13 −1 09 35 C X7 3 IR IS _3 13 −8 79 3 IR IS _3 13 −9 35 7 IR IS _3 13 −9 28 6 IR IS _3 13 −9 07 0 IR IS _3 13 −1 13 86 IR IS _3 13 −1 22 68 IR IS _3 13 −1 09 04 IR IS _3 13 −8 98 5 IR IS _3 13 −1 13 94 IR IS _3 13 −1 07 74 IR IS _3 13 −1 21 46 IR IS _3 13 −8 90 3 IR IS _3 13 −1 09 07 IR IS _3 13 −9 60 2 IR IS _3 13 −1 14 03 IR IS _3 13 −9 20 8 IR IS _3 13 −1 05 55 IR IS _3 13 −1 14 93 IR IS _3 13 −1 12 54 IR IS _3 13 −1 11 26 IR IS _3 13 −8 58 5 IR IS _3 13 −1 07 55 IR IS _3 13 −1 11 42 IR IS _3 13 −8 94 6 IR IS _3 13 −1 22 96 IR IS _3 13 −9 56 0 IR IS _3 13 −9 34 2 IR IS _3 13 −9 28 5 IR IS _3 13 −1 15 45 IR IS _3 13 −1 19 97 IR IS _3 13 −1 15 21 IR IS _3 13 −1 15 49 IR IS _3 13 −8 94 8 IR IS _3 13 −1 11 46 IR IS _3 13 −1 05 75 IR IS _3 13 −7 79 7 IR IS _3 13 −8 69 9 IR IS _3 13 −1 11 44 IR IS _3 13 −1 15 48 IR IS _3 13 −1 11 30 IR IS _3 13 −8 29 2 IR IS _3 13 −1 13 38 IR IS _3 13 −1 12 92 IR IS _3 13 −1 22 10 IR IS _3 13 −8 71 7 IR IS _3 13 −1 18 16 IR IS _3 13 −1 11 39 IR IS _3 13 −1 11 29 IR IS _3 13 −1 11 77 IR IS _3 13 −9 57 2 IR IS _3 13 −1 23 29 IR IS _3 13 −9 96 8 IR IS _3 13 −1 18 18 IR IS _3 13 −1 04 49 IR IS _3 13 −9 31 3 IR IS _3 13 −7 72 8 IR IS _3 13 −8 30 5 IR IS _3 13 −1 08 59 IR IS _3 13 −1 17 97 C X1 51 IR IS _3 13 −1 11 07 IR IS _3 13 −1 17 43 IR IS _3 13 −1 06 09 IR IS _3 13 −1 06 76 IR IS _3 13 −1 06 71 IR IS _3 13 −1 19 34 C X8 9 IR IS _3 13 −8 34 1 IR IS _3 13 −1 15 94 IR IS _3 13 −1 16 94 IR IS _3 13 −9 55 5 IR IS _3 13 −1 19 47 IR IS _3 13 −8 74 3 IR IS _3 13 −1 20 78 IR IS _3 13 −1 02 21 IR IS _3 13 −1 18 76 IR IS _3 13 −9 42 9 IR IS _3 13 −1 17 44 IR IS _3 13 −9 31 7 IR IS _3 13 −9 20 4 IR IS _3 13 −1 07 48 IR IS _3 13 −1 16 93 IR IS _3 13 −1 21 86 IR IS _3 13 −1 10 97 IR IS _3 13 −1 20 57 IR IS _3 13 −1 01 77 C X1 62 IR IS _3 13 −1 18 93 Genotype A nc es tr y co ef fic ie nt b Fig. 5   Principal component analysis of the 204 3K lines based on 9,126 high quality SNP markers. a = Scree plot showing the optimum number of ancestral populations; b = Heatmap of individual admix- ture coefficients for the optimum number of ancestral populations; c = Scatter plot illustrating genotypes groupings projected on the first two components Theoretical and Applied Genetics (2026) 139:52 52   Page 14 of 27 Fig. 6   Manhattan plots showing marker-trait associations for leaf blast (BL) and panicle blast (PB) severity. a = SSD Tropics population and b = 3K popula- tion. The red solid line repre- sents the corrected Bonferroni threshold, serving as the cutoff for significant markers Table 4   Most significant marker-trait associations within quantitative trait loci (QTL) region for leaf blast (BL) and panicle blast (PB) severity in SSD Tropics and 3K populations SNP single nucleotide polymorphism, Chr chromosome, Pos physical position on the Nipponbare reference genome Os-Nipponbare-Reference- IRGSP1.0, FA/UA favorable and unfavorable alleles, FAF favorable allele frequency, MAF minor allele frequency, − LOG10(p) = negative base 10 logarithm of P-value, PVE phenotypic variance explained, FDR false discovery rate QTL region Chr Pos (Mb) Lead SNP Trait FA/UA FAF  − LOG10(p) PVE FDR a. SSD Tropics qtl6.6 6 10.006 − 11.131 Pi2-01 BL T/A 0.09 11.30 0.07 1.0E-09 Pi2-01 PB A/T 0.09 8.25 0.02 2.3E-06 Pi2-02 BL C/G 0.09 11.50 0.06 1.0E-09 Pi2-02 PB C/G 0.09 7.98 0.02 2.3E-06 MSU7_6_10389352_A-T BL A/T 0.09 11.10 0.06 1.3E-09 MSU7_6_10389352_A-T PB A/T 0.09 8.17 0.02 8.0E-05 MSU7_6_10388389_TT-AA PB T/A 0.09 8.17 0.02 2.3E-06 MSU7_6_10388389_TT-AA BL T/A 0.09 12.10 0.06 5.2E-10 qtl8.3 8 6.269 − 7.470 chr08_6269190 BL A/T 0.37 6.74 0.19 2.4E-05 Pi33_3 BL T/G 0.36 6.17 0.19 7.6E-06 b. 3K population qtl1.6 1 6.227 − 6.320 6249810 PB A/G 0.83 6.56 0.18 7.6E-04 qtl6.18 6 10.321 − 10.389 191472769 BL G/C 0.13 7.35 0.06 4.6E-05 191432383 BL A/T 0.08 6.41 0.18 6.1E-05 qtl12.17 12 10.933 − 12.410 357138628 BL G/A 0.35 6.30 0.18 6.1E-05 Theoretical and Applied Genetics (2026) 139:52 Page 15 of 27  52 Table 5   Gene ontology and encoding proteins linked to the most significant marker-trait associations for leaf blast (BL) and panicle blast (PB) severity within SSD Tropics and 3K populations ACE1 avirulence conferring enzyme 1, Chr chromosome, Pos physical position on the Nipponbare reference genome Os-Nipponbare-Reference- IRGSP1.0, NBS-LRR nucleotide-binding site leucine-rich repeats QTL region Chr Pos (Mb) Lead GWAS-detected SNP Locus Encoding protein a. SSD Tropics qtl6.6 6 10.006 − 11.131 Pi2-01 Os06g0286700 NBS-LRR Pi2-02 Os06g0286700 NBS-LRR MSU7_6_10389352_A-T Os06g0286700 NBS-LRR MSU7_6_10388389_TT-AA Os06g0286700 NBS-LRR qtl8.3 8 6.269 − 7.470 chr08_6269190 Os08g0207500 Zinc transporter 4 Pi33_3 Pi33 ACE1-specific protein b. 3K population qtl1.6 1 6.227 − 6.320 6249810 Os01g0214300 Bromodomain protein qtl6.18 6 10.321 − 10.389 191472769 Os06g0286700 NBS-LRR 191432383 Os06g0286351 Armadillo-type fold domain qtl12.17 12 10.933 − 12.410 357138628 Os12g0294100 WD40 repeats protein SSD Tropics 3K AA AT TT 0.0 CC CG GG 1.5 3.0 4.5 6.0 7.5 9.0 0.0 1.5 3.0 4.5 6.0 7.5 9.0 Os06g0286700 B L se ve rit y SSD Tropics 3K AA AT TT 0.0 CC CG GG 1.5 3.0 4.5 6.0 7.5 9.0 0.0 1.5 3.0 4.5 6.0 7.5 9.0 Os06g0286700 SSD Tropics 0.0 GG GT TT 1.5 3.0 4.5 6.0 7.5 9.0 Pi33 B L se ve rit y a b c PB s ev er ity Fig. 7   Haplotype groups defined based on markers linked to Os06g0286700 and Pi33 for each trait in SSD Tropics and 3K popu- lations. a = leaf blast (BL) severity with Os06g0286700, b = pani- cle blast (PB) severity with Os06g0286700 and c = leaf blast (BL) severity with Pi33. Haplotypes for Os06g0286700 were identified by allele combinations of markers MSU7_6_10388389_TT-AA and 191472769 in SSD Tropics and 3K populations, respectively. Haplo- types for Pi33 were defined based on markers Pi33_3 in SSD Trop- ics population. Red point within each box represents group average disease severity Theoretical and Applied Genetics (2026) 139:52 52   Page 16 of 27 These methods performance was consistent across the two rice populations, demonstrating that GWAS-based weight- ing methods can be reliably applied in different breeding programs, particularly when targeting quantitative traits controlled by the combination of major and small effect genes like blast resistance. Here, we will discuss practical implementation of GS + GWAS model approach that maxi- mizes the use of both major-effect and small-effect loci to accelerate genetic gains for durable blast resistance in rice. Genetic diversity and population structure Substantial genetic diversity was observed in both popula- tions (Figs. 2, 3, 4 and 5). In the SSD Tropics population which consists of interconnected families sharing several common parents, the phenotypic analysis consistently revealed higher within-family variance than between- family variance across all traits (Table 2). This pattern suggests that most of the genetic variation in this popu- lation resides within families rather than between fami- lies, likely due to the segregation of alleles inherited from common parents. This observation was further supported by the molecular analysis where AMOVA results showed that within-family genetic variation (~ 124.1) was approxi- mately 8% higher than between-family variation (~ 104.0) (Table 3). This interconnected pedigree structure also con- tributed to the overlapping genetic backgrounds observed in the population structure analysis, where subpopulations were primary defined by family relationships rather than a clear genetic differentiation. In contrast, the 3K population exhibited a higher genetic diversity and differentiation, likely due to several factors, including the various variety 0.81 0.81 0.78 0.78 0.78 0.78 0.81 0.81 0.0 0.2 0.4 0.6 0.8 1.0 Unw FST−w −log10(p)−w AE−w 0.66 0.67 0.63 0.62 0.62 0.62 0.65 0.67 0.0 0.2 0.4 0.6 0.8 1.0 Unw FST−w −log10(p)−w AE−w 0.74 0.74 0.77 0.77 *** *** 0.0 0.2 0.4 0.6 0.8 1.0 Unw FST−w −log10(p)−w AE−w 0.62 0.62 0.65 0.65 0.74 0.82 0.76 *** *** 0.0 0.2 0.4 0.6 0.8 1.0 Unw FST−w −log10(p)−w AE−w MT ST 0.86 0.88 0.88 0.90 0.85 BL BL Pr ed ic tiv e ab ili ty Pr ed ic tiv e ab ili ty Pr ed ic tiv e ab ili ty Pr ed ic tiv e ab ili ty a (SSD Tropics) b (SSD Tropics) c (3K) d (3K) PB PB Fig. 8   Predictive abilities of single-trait (ST) and multi-trait (MT) weighted and unweighted (Unw) genomic prediction for each trait. a = leaf blast (BL) severity in SSD Tropics population, b = panicle blast (PB) severity in SSD Tropics population, c = leaf blast severity (BL) in 3K population and d = panicle blast (PB) severity in 3K pop- ulation. Values above each bar represent average predictive ability. FST-w = fixation index-based weighting method, − log10(p)-w = neg- ative base  10 logarithm of P value-derived weighting method, AE-w = squared additive effects-derived weighting method. *** sig- nificant at p < 0.001 Theoretical and Applied Genetics (2026) 139:52 Page 17 of 27  52 types of its accessions, such as ind1A, ind1B, indx, ind2, ind3, aro (aromatic), trop1 and admix types (Supplemen- tary file 1). The 3K population encompasses landraces, traditional cultivars, and improved lines of diverse variety types collected from different agro-ecological zones that might contribute to the observed higher genetic variability. The high broad sense heritability (0.69–0.83) and observed significant genotype-by-environment interaction highlight the contribution of both genetic and environmental factors to shaping leaf and panicle blast resistances (Table 2). Genetic architecture of leaf blast and panicle blast resistance GWAS analysis was implemented to link the observed genetic diversity to blast disease resistance and detected several marker-trait associations for leaf blast and panicle blast severity across the two populations, reinforcing blast resistance’s complex and polygenic nature (Table 4, Fig. 6). These findings are consistent with previous studies which reported several QTLs and candidate genes that contribute to blast resistance (Li et al. 2019; Tian et al. 2022; Korinsak et al. 2023; Devanna et al. 2024). In a recent meta-analysis, Devanna et al. (2024), reported 737 QTLs for blast disease from 53 independent populations, which they clustered into 71 meta-QTLs. Similarly, in a prior meta-QTL and RNA-seq analysis, Tian et al. (2022) also clustered 839 QTLs for blast resistance into 67 meta-QTLs, from which they reported more than 118 differentially expressed genes. Interestingly, distinct genetic basis of resistance for BL- and PB-specific loci in both SSD Tropics and 3K populations were identified, which suggests that tested genotypes hold different genes that respond differently to leaf and panicle infection, as reported previously (Kalia and Rathour 2019). Blast can infect the same rice plant at vegetative and reproductive 0.56 0.55 0.59 0.630.62 0.62 0.56 0.55 ** 0.0 0.2 0.4 0.6 0.8 Unw FST−w −log10(p)−w AE−w 0.94 0.74 0.95 0.78 0.99 0.78 0.94 0.75 *** *** *** *** 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Unw FST−w −log10(p)−w AE−w 0.48 0.44 0.52 0.65 0.69 0.68 0.52 0.48 *** * ** 0.0 0.2 0.4 0.6 0.8 Unw FST−w −log10(p)−w AE−w 0.84 0.53 0.80 0.75 0.93 0.78 0.86 0.57 *** *** *** 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Unw FST−w −log10(p)−w AE−w MT ST BL BL (SSD Tropics) (SSD Tropics) (3K) (3K) N or m al iz ed R M SE N or m al iz ed R M SE N or m al iz ed R M SE N or m al iz ed R M SE a b c d PB PB Fig. 9   Normalized root mean square error (RMSE) of single-trait (ST) and multi-trait (MT) weighted and unweighted (Unw) genomic prediction models. a = leaf blast (BL) severity in SSD Tropics popu- lation, b = panicle blast (PB) severity in SSD Tropics population, c = leaf blast (BL) severity in 3K population and d = panicle blast (PB) severity in 3K population. FST-w = fixation index-based weight- ing method, − log10(p)-w = negative base  10 logarithm of P value- derived weighting method, AE-w = squared additive effects-derived weighting method. *, ** and *** significant at p < 0.05, 0.01 and 0.001 Theoretical and Applied Genetics (2026) 139:52 52   Page 18 of 27 stage if disease conducive conditions prevail during the entire crop cycle, as is the case in tropical countries like Colombia. In other environments where climate conditions are less stable, rice fields face high disease pressure at vegetative or reproductive stage. Our findings highlight the detection of potential donors for leaf and panicle blast resistance under natural infection conditions. This differentiation is very useful for breeding programs since both leaf and panicle blast resistance are necessary to decrease the negative economic and environmental impact of blast disease on rice production. A major MTA, namely Pi33_33 linked to Pi33 gene harbored by QTL region qtl8.3 (6.269–7.470  Mb) was detected in SSD Tropics population for leaf blast resistance, explaining about 19% of phenotypic variance of the trait (Tables 4 and 5). Pi33 was previously reported as a major blast resistance gene, which encodes a protein that recog- nizes pathogen’s avirulence conferring enzyme 1 (ACE1) (Berruyer et al. 2003; Ballini et al. 2007). ACE1 is a M. grisea avirulence gene which encodes polyketide synthase protein fused to a non-ribosomal peptide synthetase involved in the biosynthesis of secondary metabolite that is specifi- cally recognized by Pi33 during the infection (Collemare et al. 2008). Based on gene expression analysis in the IR64 rice cultivar, Vergne et al. (2007) demonstrated that ACE1- to-Pi33 interaction triggers the activation and upregulation of defense and genes and down-regulation of several chlo- rophyll a/b binding genes, leading to enhanced resistance to blast disease. Haplotype analysis revealed 517 lines that possessed favorable allele of marker Pi33_3, showing an average leaf blast severity of 2.65. These genotypes could be incorporated in future studies focusing on marker valida- tion and development of breeder’s friendly marker for use in marker-assisted breeding strategies. In addition, our results detected three highly linked ( r2 v  ≥ 0.8) common significant markers mapped to posi- tion 10.38 Mb on chromosome 6 for BL and PB severity within SSD Tropics population. This suggests the presence of shared genetic factors governing resistance to both types of blast disease as reported by Babasaheb Aglawe et al. (2017), Noenplab et al. (2006) and Korinsak et al. (2023). This common marker was linked to the locus Os06g0286700 (10.38–10.39 Mb) which was annotated as Pi2/Pi9 gene cluster according to the Rice Annotation Project Database (RAP-DB) (Sakai et al. 2013). Based on a scale of 0–9, our results revealed that this gene reduced blast severity of best haplotypes by up to 2.4 points (Fig. 7). Kalia and Rathour (2019) found that most major blast resistance genes, includ- ing Pi2/Pi9 gene cluster encode an NBS-LRR protein. This protein is known to play a pivotal role in plant defense, particularly through pathogens’ effectors recognition and effector-triggered immunity initiation, a key plant immune response (Dubey and Singh 2018). Similarly, Devanna et al. (2024) identified NBS-LRR genes from 53 refined meta- QTLs for blast resistance, further emphasizing the impor- tance of these genes in breeding for blast resistance in rice. Moreover, our study also identified another significant marker (191472769, 10.38 Mb) linked to Pi2/Pi9 for BL severity within 3K population. This demonstrates the con- sistency of this gene across diverse genetic backgrounds and highlights its potential as a promising target for genomics- assisted breeding for improved leaf blast resistance in rice. The relatively high variability observed within haplotype groups for both BL and PB severity in the two population (Fig. 7), could be explained by the existence of several small effect genes that contribute to blast resistance besides complete and race-specific genes such as Pi2/Pi9 and Pi33 detected in this study. Furthermore, a moderate genotypic correlation (0.43–0.44) was also observed between two traits, support- ing the fact that some genetic factors influence both leaf and panicle blast resistance (Figs. 2 and 3). In this case, selection for resistance to leaf blast would positively influ- ence resistance to panicle blast disease. Similarly, Korin- sak et al. (2023), evaluated three blast isolates (THL191, THL949, and THL 557) and found moderate positive cor- relations (0.45–0.47) between leaf and panicle blast resist- ances. Noenplab et al. (2006) also reported low to moder- ate correlations (0.28–0.56) between leaf and panicle blast using various blast isolates. However, the low and moder- ate strength of the correlation also supports the presence of distinct genetic factors governing leaf and panicle blast resistance. Different correlations reported using similar data analysis are impacted by the blast population used to gener- ate phenotypic data, the results reported here correspond to natural infection where the virulence spectrum of the entire population is difficult to assess, contrary to evaluations car- ried out using purified strains. Further analysis should be done to stablish if the resistance identified in this study is able to control blast pathogen in other locations where rice blast is a major threat. G‑matrix weighting methods and genomic prediction ability The high average predictive ability of 0.74–0.78 observed for leaf blast severity and the moderate predictive ability of 0.62–0.63 for panicle blast severity with the standard unweighted single-trait GBLUP underscore the potential of GBLUP-based genomic selection to accelerate blast resist- ance breeding in rice (Fig. 8). Similar results were reported by Huang et al. (2019), who evaluated several blast isolates in two rice populations and found an average predictive abil- ity of 0.55 for GBLUP. In our study, the predictive ability was consistently high for leaf blast severity and moderate for panicle blast severity across the two populations. Average Theoretical and Applied Genetics (2026) 139:52 Page 19 of 27  52 predictive ability of unweighted GBLUP for leaf blast sever- ity compared to panicle blast severity was 0.16 and 0.12 higher in SSD Tropics and 3K populations, respectively. This may be attributed to differences in the genetic archi- tecture of these traits and their underlying molecular mech- anisms, which are likely shaped by distinct pathosystems. Despite the ability of GS to predict both traits, its efficiency might be higher for leaf blast resistance. Genetic resources for panicle blast resistance are usually limited, partly due to difficulties in phenotyping for panicle blast disease using available evaluation methods (Hayashi et al. 2019). Visual plot-wise scoring under natural infection conditions may be more reliable for leaf blast than panicle blast since foliar tissue is most exposed to the rater eye than panicle structure in which capturing moderate resistant reaction could be chal- lenging. An additional strategy as plant-wise phenotyping method could be considered to confirm the reduced associa- tion between genetic diversity and panicle resistance within the studied population and apply the models to capture all loci that may contribute to this trait. Additionally, very late genotypes may escape highest panicle blast infection pres- sure under natural conditions leading to missing records for PB severity as observed mostly in 3K population in our study (Fig. 1). Moreover, the results also confirm the ability of GBLUP to generalize across diverse genetic backgrounds and environments effectively. In addition, the application of different weighting meth- ods resulted in varying levels of gains in the predictive ability compared to the unweighted GBLUP models across populations and traits. Improvements ranged from high to no increases in the average predictive ability and were con- sistent across both ST and MT models. This demonstrates that the integration of marker-specific weights has the poten- tial to increase genomic prediction accuracy. Incorporation of marker weights accounted for differences in the genetic control of the traits as well as the biological relevance of specific genomic loci. Similar findings have been reported by Montesinos-López et al. (2024b) and Ren et al. (2021), who reported significant improvement in genomic predic- tion accuracy using marker-specific weights. Moreover, the weighting methods achieved moderate to substantial reduc- tion in the normalized root mean square error (nRMSE) compared to unweighted models (Fig. 9). These results firstly emphasize the effectiveness of weighted models in accurately ranking genotypes based on their genomic esti- mated breeding values (GEBVs), thereby improving selec- tion decisions. Secondly, they also highlight the potential of weighted models to predict GEBVs that numerically approximate the true and/or observed performance. Among the tested methods on the single-trait model, the improvement shown by the FST-w was population-specific, with no improvement observed in the average predictive ability within the SSD Tropics population for either leaf or panicle blast severity. This finding shows that population differentiation alone may not sufficiently capture the genetic architecture of blast resistance to improve genomic predic- tion accuracy. However, FST-w caused a modest increase in the predictive ability in the 3K population, ranging from 4.1 to 4.8% compared to the unweighted models. Chang et al. (2019) reported a comparable improvement and observed a 5% gain in prediction accuracy when applying FST-based weighted GBLUP in animal breeding. Nevertheless, while this method may enhance prediction accuracy for differ- ent traits, our results demonstrate that its effectiveness was inconsistent across diverse genetic backgrounds. This limitation is particularly apparent in cases of overlapping ancestral populations, where allele frequencies are similar across subpopulations, and the genetic variation between subpopulations is relatively low, as observed in the SSD Tropics population. For all traits, subpopulations in the SSD Tropics were driven by family structure, and within-family genetic variance was higher than between-family variance. In both populations, the application of AE-w and − log10(p)-w resulted in consistently higher predic- tive abilities and lowest nRMSE for all traits compared to the unweighted models and FST-w methods. Across traits and populations, the highest improvements relative to the unweighted models were 37.1% and 32.3% for AE-w and − log10(p)-w, respectively. This demonstrates the supe- riority of AE-w and − log10(p)-w as weighting methods for optimizing prediction accuracy for rice blast resistance. Marker effect-based weighting was also used by Strandén and Jenko (2024) and Ren et al. (2021), who found improve- ments in the prediction accuracy for complex traits. Moreo- ver, the results highlight the high potential of genome-wide association statistics for increasing genomic prediction accuracy through the integration of marker significance levels and additive effects into the prediction model. Su et al. (2014) also reported the superiority of P-value derived weights over other weighting methods. Similarly, Spindel et al. (2016) demonstrated that the integration of GWAS into genomic prediction using ridge regression BLUP (RR- BLUP), an alternative to GBLUP model, improved the pre- diction accuracy for several agronomic traits in rice. Zhang et al. (2023) incorporated P value-derived weight matrix into RR-BLUP model and reported a significant improve- ment in predictive ability of agronomic traits in rice. Unlike population-specific or trait-specific methods, the GWAS- based methods show broader applicability and could capture genomic signals more effectively across diverse backgrounds and trait architectures. Moreover, AE-w and − log10(p)-w effectively capture genome-wide signals associated with both large- and small-effect (polygenic background) loci. In contrast, FST-w primarily reflects population differentia- tion and may not target loci that are directly influencing the Theoretical and Applied Genetics (2026) 139:52 52   Page 20 of 27 trait of interest, which likely explains its consistently lower performance. Despite the similarity of the unweighted models’ perfor- mance between populations, particularly for panicle blast severity, our results showed that the predictive abilities with AE-w and − log10(p)-w were up to 26.9% higher in the 3K population relative to the SSD Tropics population. Across traits, the maximum within-population improve- ments achieved by AE-w and − log10(p)-w relative to the unweighted models