Genetic dissection of grain yield and correlated proxy traits under suboptimal conditions Authors: Yan-Cheng Lin, Claude Urbany, Antonina Shlykova, Armin C. Hölker, Milena Ouzunova, Thomas Presterl, Torsten Pook, Manfred Mayer, Sebastian Urzinger, Chris-Carolin Schön Published: 2026 ABSTRACT Key messageThis eight-founder MAGIC population represents a powerful resource for dissecting complex traits in maize, assessing the utility of drought proxy traits, and optimizing low-coverage whole-genome sequencing approaches.AbstractSecuring sustainable crop production requires the genetic improvement of abiotic stress tolerance. Due to the broad range of environmental factors causing abiotic stress and complex genotype-by-environment interactions, it is crucial to understand the genetic basis of crop yield under suboptimal conditions. Here, we developed a dent maize Multi-parent Advanced Generation Inter-Cross (MAGIC) population comprising 388 doubled-haploid (DH) lines. The population was derived from eight founders with varying stress tolerance, selected from a dent diversity panel evaluated for yield performance across a wide range of European environments. The MAGIC DH lines were genotyped via whole-genome sequencing (~ 5× coverage) and evaluated in seven testcross and 14 line per se trials, for grain yield, leaf senescence, leaf rolling, anthesis-silking interval, and six additional agronomic traits. Genetic dissection identified 22 grain yield QTL, explaining 45% of the genetic variance. Under heat and drought stress, testcross grain yield correlated significantly with leaf senescence and leaf rolling measured in line per se trials. Bivariate multi-trait analysis showed that alleles for delayed senescence and reduced rolling at detected QTL generally exhibited positive effects on grain yield, suggesting that accumulating these favorable alleles could enhance yield performance. Incorporating these proxies into multi-trait genomic prediction models improved yield prediction accuracy, although gains were constrained by modest trait correlations. Given the comprehensive data, we also provide recommendations for optimizing sequencing depth and QTL mapping strategies in experimental maize populations.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s00122-026-05364-w. Introduction Breeding for tolerance to water-deficit conditions has become a critical strategy to enhance resilience of our crops to climate change and ensure food security. However, directly selecting for grain yield under suboptimal conditions has remained difficult due to high variability of potential environmental scenarios and complex genotype-by-environment (G×E) interactions affecting yield performance. Indirect selection for grain yield, either by genome-based strategies or by the evaluation of proxy traits that have higher heritability, are correlated with yield, and are amenable to high-throughput phenotyping, might attenuate these limitations. During the vegetative stage, leaf rolling has been suggested as a physiological indicator of drought response in maize (Bänziger et al. 2000), but its relevance for selecting drought-tolerant genotypes has remained elusive. Saglam et al. (2014) compared two cultivars with different drought responses and reported that leaf rolling reduces water loss and mitigates stress damage by limiting transpiration in the tolerant genotype. In contrast, Effendi et al. (2019) showed that drought-tolerant varieties exhibited less leaf rolling, suggesting that excessive rolling could signal susceptibility rather than tolerance. During the reproductive stage, asynchronous development of anthesis and silking under stress has been shown to lead to severe yield loss (Bruce et al. 2002). However, modern maize hybrids have been selected for short anthesis-silking interval (Araus et al. 2012), making it less relevant when selecting for yield improvement under challenging conditions. During the grain-filling period, delayed senescence (stay-green) has been suggested as one of the mechanisms underlying stress tolerance (Tollenaar and Wu 1999). Genotypes with longer stay-green have been shown to have higher grain yield under drought-heat conditions (Cerrudo et al. 2017). However, the stay-green phenotype may hinder grain dry-down when green leaf area is retained beyond maturity (Lee and Tollenaar 2007). In a seminal study aimed at understanding the genetic control of grain yield in heat- and drought-prone environments, a diversity panel (DROPS panel) of 244 dent maize hybrids was evaluated under irrigated and rainfed conditions across nine European sites in the years 2012 and 2013 and one site in Chile in 2013, totaling 29 field experiments (Millet et al. 2016). The extensive multi-environment dataset facilitated the characterization of QTL allelic effects on yield in relation to environmental variables and supported the development of genomic prediction models that incorporated genotype-by-environment interactions (Millet et al. 2019). Building on the results of this study, we developed a Multi-parent Advanced Generation Inter-Cross (MAGIC) population using eight inbred lines selected from the diversity panel for their contrasting performance in drought- and heat-prone environments. We chose the MAGIC design due to its favorable properties for QTL mapping and genomic prediction (Scott et al. 2020). MAGIC populations are created by inter-mating several inbred founders over multiple generations before deriving homozygous inbred lines for phenotyping. This inter-mating process leads to a population with genomes that form fine-scale mosaics, reflecting balanced contributions from all founder lines (Huang et al. 2015). Compared to diversity panels, MAGIC populations have minimal population structure. Known founder genotypes and balanced allele frequencies facilitate accurate imputation when using whole-genome shallow sequencing as a genotyping platform. Compared to bi-parental populations, MAGIC populations exhibit significantly higher genomic diversity and recombination. In maize, several MAGIC populations have been developed for genetic studies. They offered high power and resolution for QTL mapping and facilitated the identification of candidate genes (Dell'Acqua et al. 2015; Ferguson et al. 2025). The genetic control of early-stage drought tolerance (Rida et al. 2021) and leaf senescence (Caicedo et al. 2021) has also been studied using MAGIC populations. Genomic prediction in MAGIC populations has been shown to have high accuracy for important agronomic traits (Michel et al. 2022). Moreover, Hudson et al. (2022) used a MAGIC population to study genotype-by-environment (G×E) interactions. Here, we evaluated an eight-founder MAGIC population across 21 diverse field experiments. Our objectives were to: (i) assess the efficiency of shallow-seq as a genotyping platform and suggest best practices, (ii) characterize the genetic and phenotypic variation of the developed MAGIC population for important agronomic traits including grain yield under suboptimal conditions, and (iii) evaluate the relevance of proxy traits for grain yield improvement. Materials and methods Genetic material A MAGIC population comprising 388 doubled-haploid (DH) lines was developed by inter-crossing eight dent maize inbred lines using three funnel crossing schemes (subfamilies), as detailed in Table S1. The eight-founder lines, B106, B107, F888, FC1890, Lo1056, Lo1270, Lo1290, and PHG83, were selected from the DROPS panel based on their grain yield performance in ten testcross trials conducted by Millet et al. (2016). These founders exhibited similar grain yield in irrigated trials, while showing variable performance in rainfed trials (Fig. S1a). Founders FC1890, B107, and PHG83 belong to the Iodent heterotic group, Lo1056, Lo1270, and Lo1290 belong to the Lancaster group, and F888 and B106 have been categorized as other Non-Stiff Stalk (NSS) material (Fig. S1b). The eight lines were selected for moderate variation in flowering time, with a maximum difference of about 8 days under irrigation and 11 days under rainfed conditions (Millet et al. 2016). Field experiments The 388 MAGIC DH lines, their founder lines, together with 18 elite inbred lines proprietary to KWS SAAT SE, were phenotyped in 21 field trials conducted between 2020 and 2023 in seven locations in Germany (4), Hungary (2), and Italy (1). Seven trials assessed testcross performance using the flint inbred line UH007 as the tester, consistent with the DROPS project, while fourteen trials evaluated line per se performance. All trials were arranged in generalized alpha-lattice designs with two replications and block sizes of either five or ten plots. In addition to the final set of 388 MAGIC DH lines, trials included DH lines that did not pass marker-based identity filtering and were treated as checks in subsequent phenotypic analyses. Two plot types were used: observation plots (3 m long, 0.75 m row spacing, planting density of 8.8 plants m−2) and yield plots (6 m long, 0.75 m row spacing, 11 plants m−2). All yield plots were planted as double-rows. Observation plots were single-row, except for two trials with double-rows in 2023. Sowing, fertilization, and plant protection followed standard agricultural practices at the respective experimental stations. Some trials were conducted as rainfed experiments on sandy soils to simulate water-limited conditions. Summary information on all field experiments is provided in Table S2. Weather data were obtained for each trial from DTN ClearAg™ (https://www.dtn.com/agriculture/agribusiness/clearag). Data covered a 60-day window spanning 30 days before and 30 days after male flowering (median across all plots) and are provided in File S1. To specify the environmental conditions associated with water and temperature stress during this period, we summarized precipitation, reference evapotranspiration (ET0, calculated using the FAO-56 Penman–Monteith equation), and maximum temperature in Fig. S2. In rainfed trials, environments in which precipitation was substantially lower than evapotranspiration, particularly in 2021 and 2022, were considered suboptimal conditions (water-stressed). Trait assessment Ten traits were measured (Table S2), including grain dry matter yield (GDY, calibrated to 85% dry matter content, t/ha; hereafter referred to as grain yield), grain dry matter content (GDC, %), early plant height (PH_V6, measured at the V6 stage from the soil surface to the tip of the tallest leaf when stretched, averaged over three plants per plot, cm), final plant height (PH_final, measured from the soil surface to the ligule of the flag leaf, averaged over three plants per plot, cm), ear height (EH, measured from the soil surface to the ligule of the upper cob, averaged over three plants per plot, cm), male flowering (MF, number of days from sowing to 50% of plants shedding pollen in a plot, days), and female flowering (FF, number of days from sowing to 50% of plants silking in a plot, days). Proxy traits included a leaf senescence score (LS, scored 4–6 weeks after flowering on a 1–9 scale, where 1 = very green and 9 = complete senescence), a leaf rolling score (LR, scored on a hot dry day on a 1–9 scale, where 1 = no leaf rolling and 9 = severe rolling across all plants), and the anthesis-silking interval (ASI, days). Leaf senescence and leaf rolling were recorded only when phenotypic variation was visible. Phenotypic data analysis The statistical model for analyzing the phenotypic data is given by:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{pijkl}={\mu}_{p}+{g}_{i\left(p\right)}+{t}_{j}+{gt}_{ij\left(p\right)}+{r}_{k\left(j\right)}+{b}_{l\left(kj\right)}+{\varepsilon}_{pijkl}$$\end{document}ypijkl=μp+gip+tj+gtijp+rkj+blkj+εpijkl Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{pijkl}$$\end{document}ypijkl is the phenotypic plot-level observation; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mu}_{p}$$\end{document}μp denotes the overall mean of population pp, where pp = 1 for MAGIC DH lines and pp = 2 for founders and checks; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g}_{i\left(p\right)}$$\end{document}gip is the effect of genotype ii nested in group pp. For variance component estimation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g}_{i\left(p\right)}$$\end{document}gip was treated as random effect for pp = 1 and as fixed for pp = 2. For obtaining adjusted entry means across trials, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g}_{i\left(p\right)}$$\end{document}gip was treated as fixed for all genotypes. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{j}$$\end{document}tj is the random effect of trial jj; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${gt}_{ij\left(p\right)}$$\end{document}gtijp is the random interaction effect of genotype ii and trial jj with heterogeneous variances \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{gt(p)}^{2}$$\end{document}σgt(p)2 among groups; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{k\left(j\right)}$$\end{document}rkj and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}_{l\left(kj\right)}$$\end{document}blkj are random effects of replication kk nested in trial jj and incomplete block ll nested in replication kk in trial jj; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon}_{pijkl}$$\end{document}εpijkl is the random residual error of observation pijklpijkl, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon}_{pijkl}\sim \mathrm{i}\mathrm{i}\mathrm{d} N\left(0,{\sigma}_{\varepsilon }^{2}\right)$$\end{document}εpijkl∼iidN0,σε2. ASReml (v4.2) (Butler et al. 2017) was used to fit linear mixed models, estimate variance components, and compute adjusted entry means. Adjusted entry means for individual trials were calculated by excluding trial (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{j}$$\end{document}tj) and genotype-by-trial (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${gt}_{ij\left(p\right)}$$\end{document}gtijp) effects from Eq. (1). Trait heritabilities and their 95% confidence intervals were estimated on an entry-mean basis following Hallauer et al. 2010 and Knapp et al. (1985), respectively. Phenotypic correlations were calculated from adjusted entry means of individual trials using Pearson correlation coefficients with false discovery rate (FDR) correction (Benjamini and Hochberg 1995) accounting for multiple testing. Whole-genome short-read sequencing The eight founders of the MAGIC population were sequenced at high depth (> 50×), while the 388 MAGIC DH lines were sequenced at an average genome sequencing depth of 5× (shallow-seq) using 150-bp paired-end short reads generated with the Illumina NovaSeq 6000 platform. For quality assessment, the eight-founder lines and a subset of 64 DH lines were also genotyped using the 600 k Affymetrix Axiom Maize Array (Unterseer et al. 2014). Initially, variant calling was performed using whole-genome sequences of the eight-founder lines. The raw reads were preprocessed by removing PCR duplicates using hts_SuperDeduper (v1.3.0) (Bonfield et al. 2021), trimming adapters and low-quality sequences with Trimmomatic (v0.39) (Bolger et al. 2014), and correcting base errors with Lighter (v1.1.2) (Song et al. 2014). Reads were aligned to the B73v5 reference genome (Hufford et al. 2021) using bwa (v0.7.17) (Li and Durbin 2009). Alignment (BAM) files were sorted using SAMtools (v1.9) (Li et al. 2009), and duplicate reads were marked with picard MarkDuplicates (v2.26.0) (https://broadinstitute.github.io/picard/). Variant calling was performed with freebayes (v1.3.5) (Garrison and Marth 2012). Variants located in low-complexity regions annotated by sdust (Morgulis et al. 2006) were removed. Bi-allelic homozygous SNPs with complete data across all founders (8.3 M) were selected as reference for the downstream variant calling in the DH lines. Reads of DH lines were preprocessed with the same pipeline for trimming, base correction, duplicate removal, and alignment. Variant calling was performed with freebayes, using the founder-derived SNP positions as reference. Reads from founder lines were jointly called as additional samples. SNP loci were excluded if founders showed inconsistent genotypes between the two variant calling runs, heterozygosity was greater than 10%, or missing rates exceeded 50% in the MAGIC population. Further filtering removed SNPs located in annotated transposable elements (Zm-B73-REFERENCE-NAM-5.0.TE.gff3.gz, downloaded from https://download.maizegdb.org/). The final dataset was imputed using the R package HBimpute (Pook et al. 2021), and SNPs with minor allele frequency (MAF) < 0.05 were discarded, yielding 2,717,240 high-quality SNPs for downstream analyses. To assess the efficiency of shallow-seq, we down-sampled reads from 31 DH lines with high depth (> 8×) to depths of 0.1×, 0.5×, 1×, 2×, 4×, and 8× using Seqkit (Shen et al. 2016). At each depth, variant calling followed the same reference SNP guided pipeline described above. Imputation was conducted on all SNP datasets without prior SNP filtering using Beagle 5.4 (Browning et al. 2018). For each sequencing depth, we evaluated four metrics: (1) genome coverage, defined as the proportion of the genome covered by reads, (2) genotyping rate, defined as the proportion of SNPs genotyped per individual, (3) number of usable SNPs, defined as SNPs with a missing rate below 50%, and (4) genotyping error rate, calculated as the discordance with the 600 k SNP array genotypes. Summary statistics were computed using SAMtools (Li et al. 2009) and VCFtools (Danecek et al. 2011). Population structure and LD decay Population structure of the DH lines was assessed using a subset of 20,974 SNPs selected from the full SNP set (2.7 M) via LD pruning (r2 < 0.4 within a 500 kb window) using PLINK (Purcell et al. 2007), balancing genome coverage and minimal LD. Pairwise genetic distances between individuals were computed using modified Rogers distances (Wright 1978). Principal Coordinate Analysis (PCoA) was conducted using the pcoa() function in R. A neighbor-joining phylogenetic tree (Saitou and Nei 1987) was constructed utilizing the nj() function from package ape (Paradis and Schliep 2019) and visualized with the phytools (Revell 2012) package. For LD decay analysis, the full SNP dataset was thinned by selecting every 10th SNP to reduce computational demands. Pairwise LD (r2) between SNPs was computed using PopLDdecay (Zhang et al. 2019). LD decay distance was estimated for an r2 threshold of 0.2 (Hill and Weir 1988). Genomic mosaic reconstruction and recombination estimation Founder haplotype (FH) probability, i.e., the likelihood that a specific allele in a progeny line originated from a given founder, was estimated using the R package HaploBlocker (v1.7.1) (Pook et al. 2019) with function founder_detection(), which employs a Hidden Markov Model (HMM), similar to the approach used in mpmap2 (Shah et al. 2019). HaploBlocker improves upon mpmap2 by using a haplotype library constructed from both founders and progeny, which adjusts emission probabilities when shared haplotype blocks are detected. To reduce computational load and minimize noise from genotyping errors, the 2.7 million SNP dataset was partitioned into 50 subsets by assigning every 50th marker to the same subset. Founder probabilities were computed independently for each subset and subsequently averaged across a 50-marker sliding window. While this approach entails a reduction in mapping resolution, it significantly mitigates the influence of consecutive genotyping errors. Founders may share chromosome segments that are identical by descent (IBD). Pairwise IBD regions of founders were detected using phasedibd (Freyman et al. 2021) with default settings, and IBD proportions were estimated based on the length of IBD regions on the linkage map. As these segments likely reflect common ancestry, founder alleles were merged into an ancestral allele if the corresponding chromosome regions were IBD between founders. For each DH line, ancestral haplotype (AH) probabilities were obtained by summing founder haplotype probabilities within each ancestral group. Recombination event counts were estimated using the haploRILs R package (https://github.com/GoliczGenomeLab/haploRILs). The method compares the genomic composition of each DH line to the founder genomes by assigning founder haplotype blocks along each chromosome. The SNP dataset was divided into 50 subsets (every 50th marker), and estimations were performed independently for each subset using the settings nSnp = 20, step = 1, and k = 1. Final recombination counts per chromosome per individual were calculated as the average across all 50 runs. Genome-wide association (GWA) analysis in individual trials We first performed a GWA analysis on 388 DH lines using bi-allelic SNPs. The dataset was pruned with PLINK (Purcell et al. 2007) to remove redundant markers (r2 > 0.99 within a 300 kb window). Given the high marker density, SNPs that showed r2 < 0.8 with all other markers within 10 Mb distance were considered erroneous and excluded. This resulted in a final set of 359,141 SNPs. Association analyses were conducted on the adjusted entry means from individual trials and means across trials using vcf2gwas (Vogt et al. 2022), which implements a mixed linear model of GEMMA (Zhou and Stephens 2012):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{y}=\mathbf{1}\mu +\mathbf{x}\beta +\mathbf{Z}\mathbf{u}+{\boldsymbol{\upvarepsilon}}$$\end{document}y=1μ+xβ+Zu+εwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{y}$$\end{document}y is the nn-dimensional vector of adjusted entry means, with nn being the number of MAGIC DH lines (nn = 388); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{1}$$\end{document}1 is a vector of ones; µ is the overall mean; ββ is the fixed effect of the tested SNP marker; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{x}$$\end{document}x is the vector of corresponding SNP genotype scores coded as 0 and 2; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{u}$$\end{document}u is the nn-dimensional vector of random polygenic effects, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{u}\sim N\left(0,\mathbf{K}{\sigma}_{g}^{2}\right)$$\end{document}u∼N0,Kσg2, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2 being the genetic variance pertaining to the model. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{K}$$\end{document}K denotes the n× nn×n realized genomic relationship matrix, computed using the method of Astle and Balding (2009) implemented with the leave-one-chromosome-out (LOCO) approach described by Yang et al. (2014); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{Z}$$\end{document}Z is the design matrix for the polygenic effects in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{u}$$\end{document}u; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\upvarepsilon}}$$\end{document}ε is the nn-dimensional vector of random residual effects, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\upvarepsilon}}\sim N\left(0,{\mathbf{I}}_{\mathbf{n}}{\sigma }^{2}\right)$$\end{document}ε∼N0,Inσ2 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{I}}_{\mathrm{n}}$$\end{document}In being an n× nn×n identity matrix and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }^{2}$$\end{document}σ2 the residual variance. SNP-trait associations were assessed using the Wald test (Wald 1943), and significance was determined at a 1% false discovery rate (FDR) (Benjamini and Hochberg 1995). Significant SNPs with r2 > 0.2 within 3 Mb distance were merged into a single QTL, with the SNP showing the lowest p-value designated as the lead SNP. All lead SNPs were fitted in a multi-locus model using stepwise backward elimination, with each SNP tested as last fixed term using the Wald test. SNPs with p-values ≥ 0.01 were iteratively removed until all remaining SNPs met the threshold, resulting in a specific QTL set (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{Q}}$$\end{document}Q) for each trait in each trial. QTL heritability, i.e., the proportion of phenotypic variance explained by individual QTL, was estimated by treating QTL effects as random and using centered and scaled genotype coding to calculate their variance relative to the total phenotypic variance, following Wang et al. (2022) with the following model:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{h}}_{QTL}^{2}=\frac{{\widehat{\sigma }}_{q}^{2}}{\sum_{q\in {\boldsymbol{Q}}}{\widehat{\sigma }}_{q}^{2}+{\widehat{\sigma }}_{g}^{2}+{\widehat{\sigma }}^{2}}$$\end{document}h^QTL2=σ^q2∑q∈Qσ^q2+σ^g2+σ^2where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{\sigma }}_{q}^{2}$$\end{document}σ^q2 is the variance estimate of a particular QTL; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{Q}}$$\end{document}Q denotes the QTL set of a trait in a trial; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{\sigma }}_{g}^{2}$$\end{document}σ^g2 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{\sigma }}^{2}$$\end{document}σ^2 are the variance estimates of polygenic effects and residuals, respectively. Multi-trial model for final QTL set To consolidate QTL for each trait, we performed stepwise backward elimination on all QTL detected from individual trial analyses, adapting the multi-locus, multi-trial model from Millet et al. (2016):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{ij}=\mu +{t}_{j}+\sum_{q\in {{\boldsymbol{Q}}}^{\boldsymbol{*}}}{x}_{q}{\beta}_{q}^{j}+{u}_{i}+{\varepsilon}_{ij}$$\end{document}yij=μ+tj+∑q∈Q∗xqβqj+ui+εijwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{ij}$$\end{document}yij is the adjusted mean of genotype ii in trial jj; μμ is the overall mean; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{j}$$\end{document}tj is the fixed effect of trial jj; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{q}$$\end{document}xq is the SNP genotype score (0 or 2); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta}_{q}^{j}$$\end{document}βqj is the fixed effect of SNP qq in trial jj; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{i}$$\end{document}ui is the random polygenic effect of genotype ii. The kinship matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{K}}^{\boldsymbol{*}}$$\end{document}K∗ was calculated based on all chromosomes as multiple QTL are fitted in the model; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon}_{ij}$$\end{document}εij is the random residual error with trial-specific error variance. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{Q}}}^{\boldsymbol{*}}$$\end{document}Q∗ denotes the final QTL set. To avoid collinearity, nearby QTL (r2 > 0.8 and < 3 Mb) from different trials were grouped retaining only the marker with the lowest p-value for backward elimination. Stepwise backward elimination was then performed as described above to obtain the final QTL set (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{Q}}}^{\boldsymbol{*}}$$\end{document}Q∗). The allelic effects of the QTL remaining in the model were decomposed into a main effect and a QTL-by-trial (QxT) interaction (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta}_{q}^{j}={\upalpha}_{q}+{\upalpha}_{q}^{j}$$\end{document}βqj=αq+αqj), and Wald tests of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\upalpha}_{q}^{j}$$\end{document}αqj were used to identify consistent or trial-specific QTL (p < 0.01). Given the heterogeneous residual variances in the multi-trial analyses, we estimated the proportion of total genetic variance captured by all QTL in the final set. QTL effects were treated as fixed, and the proportion was calculated based on the reduction in genetic variance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{\sigma }}_{g}^{2}$$\end{document}σ^g2) comparing models with and without QTL. QTL confidence intervals were defined by the two most distant SNP markers in strong LD with the lead SNP (r2 > 0.8). Allelic effects at the lead SNPs were obtained as best linear unbiased estimates (BLUEs) of ββ in Eq. (2) in individual trials and Eq. (4) for the final QTL set across trials. A flowchart of the full QTL analysis is available in Fig. S3. QTL nomenclature QTL nomenclature was adapted from the convention proposed by McCouch (1997); for example, qGDY(LP)01A, where "GDY" refers to the trait (grain dry matter yield), "LP" in parentheses indicates the line per se trial (TC for testcross), "01" indicates chromosome 1, and "A" denotes the first QTL on that chromosome based on its physical position. IBD-based QTL mapping Odell et al. (2022) highlighted the value of multiple mapping approaches for QTL detection in a MAGIC population. Therefore, we also used founder (FH) and ancestral haplotypes (AH) as genomic predictors for QTL analyses. For individual trials, QTL detection was performed using an IBD-based mixed linear model implemented in the R package statgenMPP (Li et al. 2022), in which the effects of founder or ancestral alleles at a putative QTL were modeled as random effects for hypothesis testing. On the multi-trial analysis, a backward QTL elimination procedure was applied, adapting the multi-allelic, multi-locus hypothesis testing framework of Li et al. (2024). Detailed methodological procedures are described in Method S1. Multi-trait GWA analysis We studied potential pleiotropy of QTL for proxy traits and grain yield using bivariate multi-trait models. Each model included adjusted means of grain yield estimated across testcross trials together with adjusted means of one proxy trait estimated across line per se trials. The bivariate multi-trait model of Korte et al. (2012) was implemented using the R package StatgenQTLxT (https://github.com/Biometris/statgenQTLxT). Extending the model in Eq. (2) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{y}$$\end{document}y is an 2n× 12n×1 vector (nn = 388) stacking the centered and scaled phenotypes of grain yield and the proxy trait; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\upbeta}}$$\end{document}β is a 2× 12×1 vector of fixed SNP effects; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{X}$$\end{document}X is a 2n× 22n×2 design matrix for fixed effects, constructed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{I}}_{2}\otimes \mathbf{x}$$\end{document}I2⊗x, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{x}$$\end{document}x is the n× 1n×1 vector of SNP genotype scores (coded as 0 and 2); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{u}$$\end{document}u is the 2n× 12n×1 vector of random polygenic effects, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{u}\sim N\left(0, {{\boldsymbol{\Sigma}}}_{\mathbf{g}}\otimes \mathbf{K}\right)$$\end{document}u∼N0,Σg⊗K, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Sigma}}}_{\mathbf{g}}$$\end{document}Σg is a 2× 22×2 genetic variance–covariance matrix between traits, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{K}$$\end{document}K is the same n× nn×n kinship matrix used in Eq. (2), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{Z}$$\end{document}Z is the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{I}}_{2\mathbf{n}}$$\end{document}I2n design matrix, for the polygenic effects \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{u}$$\end{document}u; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\upvarepsilon}}$$\end{document}ε is the 2n× 12n×1 vector of residuals, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\upvarepsilon}}\sim N\left(0,{{\boldsymbol{\Sigma}}}_{{\boldsymbol{\upvarepsilon}}}\otimes {\mathbf{I}}_{\mathbf{n}}\right)$$\end{document}ε∼N0,Σε⊗In, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Sigma}}}_{\varepsilon }$$\end{document}Σε is the 2× 22×2 residual variance–covariance matrix between traits. Both \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Sigma}}}_{\mathbf{g}}$$\end{document}Σg and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\Sigma}}}_{{\boldsymbol{\upvarepsilon}}}$$\end{document}Σε were modeled as unstructured matrices. IBS- and IBD-based genomic prediction Genomic best linear unbiased prediction (GBLUP) was used to obtain the genetic values of the 388 DH lines. The model was a modification of Eq. (2) excluding the fixed marker effect. The kinship matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{K}$$\end{document}K was replaced by a SNP-based genomic relationship matrix (GRM), denoted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{U}}_{\mathbf{S}}$$\end{document}US, calculated following method 1 described by VanRaden (2008). The GBLUP model was implemented using ASReml (v4.2) (Butler et al. 2017). In addition to the SNP-based GRM, we constructed GRMs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{U}}_{\mathbf{F}\mathbf{H}}$$\end{document}UFH and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{U}}_{\mathbf{A}\mathbf{H}}$$\end{document}UAH using founder and ancestral haplotype probabilities. At each marker locus, each DH line was assigned a vector indicating the probability of carrying a specific founder or ancestral allele. The probability that two DH lines shared the same founder allele at a locus was calculated as the dot product of their probability vectors. The GRM was then calculated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{U}=\frac{\sum_{l=1}^{L}{(\mathbf{M}{\mathbf{M}}^{\mathbf{t}})}^{(l)}}{L}$$\end{document}U=∑l=1L(MMt)(l)L, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{M}$$\end{document}M is the n× pn×p probability matrix (nn = number of DH lines, pp = number of haplotypes) at a marker locus, and the GRM is average across all LL marker loci. For constructing the three GRMs, we used the same evenly spaced 20 K marker set described in Method S1. To test if the three GRMs differed significantly, we performed a Mantel test (Mantel 1967). Multi-trait genomic prediction A multivariate linear mixed model was implemented in ASReml (v4.2) (Butler et al. 2017) to perform multi-trait GBLUP. We modified the multi-trait GWAS model by removing the fixed marker effect and replacing the kinship matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{K}$$\end{document}K) by the SNP-based GRM (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{U}}_{\mathbf{S}}$$\end{document}US) between DH lines. Cross-validation To compare the prediction accuracy of different models and traits, we performed cross-validation. In all cross-validations, the training set comprised 250 DH lines, and the remaining DH lines were assigned to the prediction set. Each cross-validation consisted of 100 random iterations. Prediction accuracy was calculated as the correlation between predicted genetic values and observed phenotypic values in the prediction set, divided by the square root of the trait heritability. For each multi-trait model, we compared two cross-validation scenarios: (1) MT-CV1, multi-trait models in which GDY from individual trials (target trait) and adjusted means across line per se trials of leaf rolling and leaf senescence (secondary traits) were included in the training set; and (2) MT-CV2, where phenotypes of the secondary traits were additionally provided in the prediction set. Results Sequencing-based genotyping Shallow-seq of the 388 MAGIC DH lines resulted in an average depth of 5.7×. A total of 7.6 M SNPs with a missing rate less than 50% were retained from the 8.3 M SNPs identified in the eight-founder lines. After filtering and imputation, a final SNP dataset containing 2,717,240 SNPs and no missing data was obtained for downstream analyses. Using the 600 k array data as reference, the average genotyping error was estimated at 0.08%, comparable to the 0.04% observed in deep sequencing (50 ×) of the founder lines. We assessed the genotyping efficiency of the shallow-seq by down-sampling the reads to six different depths (0.1–8 ×) using 31 DH lines (Fig. 1). Genotyping rate and genome coverage increased logarithmically with sequencing depth. Variance across samples was negligible. After filtering SNPs with 50% missing rate, SNP count rose sharply from 1× to 2× depth, with 6.2 M of the total 8.3 M SNPs retained at 2× (Fig. 1a). The genotyping error rate at positions supported by reads (unimputed dataset) remained consistent across depths. In contrast, the error rate of the imputed SNP dataset decreased drastically as depth increased, with 2× or higher achieving the same error rate as the raw dataset in which all SNP loci were supported by reads (Fig. 1b).Fig. 1Efficiency and accuracy of shallow-seq. In both panels, the x-axis represents sequencing depth (0.1–8 ×). a Genotyping efficiency: mean genome coverage (red), mean genotyping rate (blue), and the proportion of SNPs with < 0.5 missingness relative to the full set of 8.3 M SNPs (green). Numbers above the green points indicate the count of retained SNPs in thousands (k). b Genotyping error rate (%) for unimputed (brown) and imputed (green) SNP datasets, with standard deviation. Error bars represent standard deviations in both panels Efficiency and accuracy of shallow-seq. In both panels, the x-axis represents sequencing depth (0.1–8 ×). a Genotyping efficiency: mean genome coverage (red), mean genotyping rate (blue), and the proportion of SNPs with < 0.5 missingness relative to the full set of 8.3 M SNPs (green). Numbers above the green points indicate the count of retained SNPs in thousands (k). b Genotyping error rate (%) for unimputed (brown) and imputed (green) SNP datasets, with standard deviation. Error bars represent standard deviations in both panels Genomic characterization of the MAGIC population The balanced intercrossing design minimized population structure in the MAGIC population. The first-principal coordinate (PCo1) explained only 2.3% of the variance, with no obvious clustering among the three subfamilies. A neighbor-joining (NJ) phylogenetic tree showed that the DH lines were distributed at equal distances from the root. LD decay ranged from 1.2 Mb to 3.3 Mb (r2 = 0.2), with chromosome 10 exhibiting significantly longer LD compared to other chromosomes (Fig. S4). Using the 2.7 M shallow-seq-based SNP markers, we reconstructed the mosaic of founder genomes in each MAGIC DH line. Figure 2 illustrates the relative contribution of founder alleles along the ten chromosomes. The contributions of the eight founders are close to the expected value of 12.5%, consistent with balanced crossing. Along most chromosomal regions, founder contributions varied mildly. However, an unexpected pattern was observed on chromosome 9. Upon closer inspection, this pattern most likely resulted from the residual heterozygosity of inbred founder Lo1056. Thus, the plant sequenced to represent this founder genotype was polymorphic in a small region of chromosome 9 compared to those used for population development. Since this pattern was confined to a specific region, the material was retained for analysis.Fig. 2a Contributions of the eight-founder lines to the MAGIC population along the 10 chromosomes. b Proportion of founder genome in the 388 DH lines a Contributions of the eight-founder lines to the MAGIC population along the 10 chromosomes. b Proportion of founder genome in the 388 DH lines On average, 3.8 recombination events were detected per chromosome, equivalent to 1.3 recombination events per meiosis (Fig. S5a). The estimation represents a lower bound as some recombination events occurring within IBD regions shared between founders or double crossovers remain undetected. The proportion of the genome in IBD between founder pairs ranged from 1 to 64%, with the maximum observed between Lo1056 and Lo1270, congruent with pedigree information that Lo1056 is the parent of Lo1270 (Fig. S5b). Due to the IBD regions between founders, the assignment of founder alleles remained ambiguous in some genomic regions. On average, 31% of the genome across the 388 DH lines could not be assigned to any founder when applying a threshold of 0.9 on founder probability. By merging founder alleles into ancestral alleles, we observed an average of 5.9 ancestral alleles per locus as compared to eight-founder alleles. Summing up the probabilities of the corresponding founder alleles to assign each ancestral allele reduced the unassigned proportion of the genome to 0.8%. Phenotypic characterization of the MAGIC population Grain yield was measured in seven testcross trials and four line per se trials with average yield ranging from 4.8 to 10.1 t/ha and 1.3 to 3.8 t/ha, respectively (Fig. 3). In 2020, high rainfall at the test sites resulted in high yield in both irrigated and rainfed trials, representing optimal conditions. In 2021, high temperature and severe water stress occurred in Hungary clearly differentiating the two water regimes in Murony (MUR). The two German rainfed test sites (DTH and PNF) showed low to intermediate performance for grain yield. In 2022, line per se trials in Hungary and at one German site experienced stress conditions (MUR and DTH) leading to highly skewed yield distributions while growing conditions at the other two German sites (ROG and MER) were optimal for maize.Fig. 3Density distribution of grain yield (GDY) based on adjusted entry means of 388 MAGIC DH lines evaluated as testcrosses (TC, upper panel) and lines per se (LP, lower panel) in seven (TC) and four (LP) trials. The vertical dashed lines indicate population means. Colors denote years: blue (2020), green (2021), and red (2022). Trial legends denote location, year, management (I = irrigated and R = rainfed), and yield rank in brackets Density distribution of grain yield (GDY) based on adjusted entry means of 388 MAGIC DH lines evaluated as testcrosses (TC, upper panel) and lines per se (LP, lower panel) in seven (TC) and four (LP) trials. The vertical dashed lines indicate population means. Colors denote years: blue (2020), green (2021), and red (2022). Trial legends denote location, year, management (I = irrigated and R = rainfed), and yield rank in brackets As summarized in Table S3, pronounced genotype-by-trial interactions were observed for grain yield, with the G×T variance component (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{gt}^{2}$$\end{document}σgt2) slightly exceeding the genetic component (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2) with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{gt}^{2}:{\sigma}_{g}^{2}$$\end{document}σgt2:σg2 ratios of 1.03 and 1.14 in testcross and line per se trials, respectively. Among the three proxy traits in line per se trials, leaf senescence showed the highest G×T ratio (1.02), followed by leaf rolling (0.78) and anthesis-silking interval (0.42). For testcross trials, the ratios were 0.88 for leaf senescence (based on two trials) and 0.72 for anthesis-silking interval, while leaf rolling was only scored in a single trial. Agronomic traits such as final plant height and flowering time showed ratios below 0.3 for both testcross and line per se trials. Pairwise phenotypic correlations among trials are shown in Fig. S6. Correlations for grain yield ranged from 0.06 to 0.79 among testcross trials and from 0.06 to 0.54 among line per se trials. As expected, correlations between testcross and line per se trials were generally low and not significant for grain yield. The three proxy traits displayed moderate correlations between trials, and significant correlations were observed between testcrosses and line per se performance. The agronomic traits final plant height and female flowering time showed high and consistent correlations across trials and between testcrosses and lines per se. Trait correlations between grain yield and the three proxy traits The trials in Murony in 2021, where high temperatures and water stress occurred (Fig. S2), allow a closer look at correlations between grain yield and proxy traits with rainfed and irrigated trials side by side. Testcross grain yield decreased by 32% in the rainfed trial compared to the irrigated trial. In the rainfed trial, grain yield showed significant correlations with both leaf senescence (− 0.25) and leaf rolling (− 0.27), suggesting that delayed senescence and reduced leaf rolling were associated with higher yield (Fig. S7a). In particular, the top 10% genotypes exhibited greater stay-green (delayed senescence) and minimal leaf rolling than the remaining lines (Fig. S7b). In contrast, the anthesis-silking interval showed no significant correlation with grain yield. We also assessed the correlations of the proxy traits evaluated on a line per se basis with testcross grain yield (Table 1). Leaf senescence measured across line per se rainfed trials showed negative correlations with testcross grain yield, reaching significance in trials at Murony (MUR), where temperatures were elevated. Leaf rolling followed a similar pattern but with weaker correlations. In contrast, anthesis-silking interval showed no consistent trend across testcross trials. Table 1Pearson’s correlation coefficients between grain yield performance from seven testcross trials (and across trials) and the three proxy traits, leaf senescence (LS), leaf rolling (LR), and anthesis-silking interval (ASI), averaged across all line per se rainfed trials20202021–KAB-RKAB-IMUR-IMUR-RMUR-IDTH-RPNF-RAcrossLS− 0.10− 0.04− 0.17**− 0.21**− 0.18**− 0.070.05− 0.13*LR− 0.06− 0.04− 0.05− 0.16*− 0.20**− 0.08− 0.10− 0.11ASI0.15**0.13*0.060.070.05− 0.04− 0.130.09Asterisks indicate significance levels after multiple testing adjustment: p < 0.05 (*) and p < 0.01 (**) Table 1Pearson’s correlation coefficients between grain yield performance from seven testcross trials (and across trials) and the three proxy traits, leaf senescence (LS), leaf rolling (LR), and anthesis-silking interval (ASI), averaged across all line per se rainfed trials20202021–KAB-RKAB-IMUR-IMUR-RMUR-IDTH-RPNF-RAcrossLS− 0.10− 0.04− 0.17**− 0.21**− 0.18**− 0.070.05− 0.13*LR− 0.06− 0.04− 0.05− 0.16*− 0.20**− 0.08− 0.10− 0.11ASI0.15**0.13*0.060.070.05− 0.04− 0.130.09Asterisks indicate significance levels after multiple testing adjustment: p < 0.05 (*) and p < 0.01 (**) Pearson’s correlation coefficients between grain yield performance from seven testcross trials (and across trials) and the three proxy traits, leaf senescence (LS), leaf rolling (LR), and anthesis-silking interval (ASI), averaged across all line per se rainfed trials Asterisks indicate significance levels after multiple testing adjustment: p < 0.05 (*) and p < 0.01 (**) Overview of QTL detection GWA analyses were conducted on adjusted entry means from individual trials and across trials. Table 2 summarizes the QTL sets obtained for grain yield and the three proxy traits in multi-trial analyses (see Table S4 for all traits). For testcross yield, 22 QTL were retained out of 71 single-trial associations, together explaining 45% of the genetic variance. Founder (FH) and ancestral (AH) haplotype-based mapping detected ten and seven QTL, respectively, with six and five overlapping with QTL regions detected by SNPs (see File S2 for the complete QTL list). For line per se grain yield, no QTL was detected using SNPs, while FH and AH methods identified 1 and 4 QTL, with no overlap between them. Table 2Summary of QTL for grain dry mater yield (GDY) and the three proxy traits, leaf senescence (LS), leaf rolling (LR), and anthesis-silking interval (ASI). nTrial refers to the number of trialsnTrialSNPFHAH\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2 explainedTestcrossGDY722 (71)10 (24)7 (22)0.45LS213 (20)5 (6)5 (6)0.45*LR1− (5)− (2)− (2)− (0.50)ASI79 (13)2 (3)2 (4)0.53Line per seGDY40 (0)1 (1)4 (6)0LS624 (36)5 (9)4 (7)0.53LR522 (47)7 (17)7 (16)0.30ASI1315 (77)3 (26)1 (27)0.70Columns SNP, FH, and AH indicate the number of QTL identified by each method. Numbers in parentheses represent total number of QTL detected in individual trials. The proportion of genetic variance explained by all QTL identified using the SNP method is given by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2 explained*Testcross leaf rolling was measured in a single trial; therefore, only individual trial QTL counts are reported Table 2Summary of QTL for grain dry mater yield (GDY) and the three proxy traits, leaf senescence (LS), leaf rolling (LR), and anthesis-silking interval (ASI). nTrial refers to the number of trialsnTrialSNPFHAH\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2 explainedTestcrossGDY722 (71)10 (24)7 (22)0.45LS213 (20)5 (6)5 (6)0.45*LR1− (5)− (2)− (2)− (0.50)ASI79 (13)2 (3)2 (4)0.53Line per seGDY40 (0)1 (1)4 (6)0LS624 (36)5 (9)4 (7)0.53LR522 (47)7 (17)7 (16)0.30ASI1315 (77)3 (26)1 (27)0.70Columns SNP, FH, and AH indicate the number of QTL identified by each method. Numbers in parentheses represent total number of QTL detected in individual trials. The proportion of genetic variance explained by all QTL identified using the SNP method is given by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2 explained*Testcross leaf rolling was measured in a single trial; therefore, only individual trial QTL counts are reported Summary of QTL for grain dry mater yield (GDY) and the three proxy traits, leaf senescence (LS), leaf rolling (LR), and anthesis-silking interval (ASI). nTrial refers to the number of trials Columns SNP, FH, and AH indicate the number of QTL identified by each method. Numbers in parentheses represent total number of QTL detected in individual trials. The proportion of genetic variance explained by all QTL identified using the SNP method is given by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma}_{g}^{2}$$\end{document}σg2 explained *Testcross leaf rolling was measured in a single trial; therefore, only individual trial QTL counts are reported Figure 4 presents the allelic effects of the 22 QTL discovered for testcross grain yield. Five QTL had consistent effects across all seven trials, while 17 QTL exhibited significant Q × T interaction. The main effects at the consistent QTL ranged from 0.27 to 0.42 t/ha, with qGDY(TC)08A on chromosome 8 having the strongest effect. This QTL was also identified by both IBD-based methods, showing that PHG83 and Lo1290 carried the favorable alleles for yield improvement (Fig. S8a). For QTL with significant QT interaction, response patterns varied. For instance, the QTL on chromosomes 2 (02A) and 7 (07A) displayed opposite effects under optimal and stress conditions; the QTL on chromosome 1 (01A) had no effect under stress but under optimal conditions, whereas the QTL on chromosome 9 (09B) had a pronounced effect mainly in trials at Murony 2021, which experienced strong heat and drought stress.Fig. 4Grain yield QTL effects across individual testcross trials (upper panel) and combined main effect (lower panel). The x-axis shows the QTL position on the linkage map, and the y-axis displays individual trials ordered by their average grain yield (GDY). Dot color indicates the direction of the allelic effect when the reference allele is substituted with the alternative allele. Solid or open circles differentiate significant from non-significant QTL in a given trial. Circle size reflects the magnitude of the QTL effect in that trial Grain yield QTL effects across individual testcross trials (upper panel) and combined main effect (lower panel). The x-axis shows the QTL position on the linkage map, and the y-axis displays individual trials ordered by their average grain yield (GDY). Dot color indicates the direction of the allelic effect when the reference allele is substituted with the alternative allele. Solid or open circles differentiate significant from non-significant QTL in a given trial. Circle size reflects the magnitude of the QTL effect in that trial For proxy traits, 13 QTL were detected for leaf senescence and seven for anthesis-silking interval, explaining 45 and 53% of the genetic variance in testcross trials, respectively. Five QTL for leaf rolling were identified from a single testcross trial, accounting for 50% of the genetic variance (Table 2). In line per se trials, 24, 22, and 15 QTL were identified for the three traits, explaining 53, 30, and 70% of the genetic variance, respectively. Among the proxy trait QTL in line per se trials, four leaf senescence, two leaf rolling, and four anthesis-silking QTL overlapped with grain yield QTL intervals. However, the lead SNPs for proxy traits and grain yield were not in tight LD (maximum r2 = 0.59), suggesting that they are unlikely to be the same QTL with pleiotropic effects. We fitted a bivariate multi-trait model incorporating grain yield (across testcross trials) and one of the proxy traits (across line per se trials) to jointly estimate SNP effects for both traits. For the lead SNPs of QTL detected for leaf senescence and leaf rolling, effect estimates were significantly correlated with effects for grain yield at the same marker position (Fig. 5). This was not the case for the anthesis-silking interval QTL. A similar pattern was observed for the lead SNPs of the grain yield QTL (Fig. S9). Aligning with phenotypic correlations, alleles reducing leaf senescence and leaf rolling generally increased grain yield, and vice versa. Notably, QTL qLS(LP)01C (leaf senescence) and qLR(LP)03C (leaf rolling) displayed consistent effects across trials and had a substantial influence on testcross grain yield (Fig. 5), marking them as promising targets for further investigation.Fig. 5Correlation of SNP effects between grain yield and proxy traits at genomic positions of lead SNPs identified for proxy trait QTL. The y-axis shows SNP effect estimates on grain yield across all testcross trials, while the x-axis shows effects of the same SNP on a leaf senescence and b leaf rolling and c anthesis-silking interval across all line per se trials. Pearson correlation coefficients and associated significance levels are shown in each panel; dot colors indicate whether the QTL exhibits significant QTL-by-Trial (Q×T) interaction. Two environment-consistent QTL (non-significant Q×T) with substantial effects on grain yield, qLS(LP)01C and qLR(LP)03C, are highlighted Correlation of SNP effects between grain yield and proxy traits at genomic positions of lead SNPs identified for proxy trait QTL. The y-axis shows SNP effect estimates on grain yield across all testcross trials, while the x-axis shows effects of the same SNP on a leaf senescence and b leaf rolling and c anthesis-silking interval across all line per se trials. Pearson correlation coefficients and associated significance levels are shown in each panel; dot colors indicate whether the QTL exhibits significant QTL-by-Trial (Q×T) interaction. Two environment-consistent QTL (non-significant Q×T) with substantial effects on grain yield, qLS(LP)01C and qLR(LP)03C, are highlighted Prediction accuracies for trait performance across trials We evaluated accuracies of whole-genome-based prediction for grain yield, the three proxy traits, and agronomic traits in the MAGIC population using SNP-based GBLUP (Table S5). Accuracies ranged from 0.33 (LS) to 0.74 (GDC) in testcross trials and 0.40 (LS) to 0.61 (FF) in line per se trials. Testcross yield could be predicted with intermediate accuracy (0.60). Among the three proxy traits measured in line per se trials, leaf rolling had the highest accuracy (0.60), followed by anthesis-silking interval (0.52) and leaf senescence (0.40). We further investigated whether constructing genomic relationship matrices using IBD information (FH and AH) could enhance prediction accuracy. Elements of the matrices derived from FH and AH probabilities were significantly correlated (r = 0.91 and 0.94, respectively) with those of the SNP-based GRM. In terms of prediction accuracy, the AH method performed as well as the SNP method, with differences in accuracy ranging from − 0.02 to 0.01. The FH method was slightly inferior, with only two of 18 traits exceeding accuracies of the SNP method, and differences ranging from − 0.04 to 0.01 (Table S5). Genomic prediction using leaf senescence and leaf rolling as secondary traits We compared single-trait and multi-trait models to evaluate line per se leaf senescence and leaf rolling performance as secondary traits for genomic prediction of testcross grain yield. Using both traits as secondary traits in the MT-CV1 model (secondary traits available only for the training set) gave similar accuracies for prediction of testcross yield as the single-trait model. In MT-CV2 (secondary traits available in both training and prediction sets), slight improvements were observed in Murony 2021 trials, from 0.53 to 0.56 under rainfed and 0.47 to 0.51 under irrigated conditions. Accuracies in the other trials remained similar to the single-trait model. Analyzing the two proxy traits separately in MT-CV2 suggested that leaf senescence was the main driver of the observed improvement in Murony (Fig. 6).Fig. 6Comparison of grain yield prediction accuracies between single-trait (ST) and multi-trait (MT) models. MT models were evaluated when secondary traits were available only in the training set (CV1) or in both training and prediction sets (CV2). The x-axis shows ST prediction accuracies, and the y-axis shows MT prediction accuracies: a MT-CV1 using both leaf senescence (LS) and leaf rolling (LR) as secondary traits; b MT-CV2 using both LS and LR; c MT-CV2 using LS only; and d MT-CV2 using LR only. Colors denote seven individual trials and across trials. Values represent means of 100 cross-validation runs Comparison of grain yield prediction accuracies between single-trait (ST) and multi-trait (MT) models. MT models were evaluated when secondary traits were available only in the training set (CV1) or in both training and prediction sets (CV2). The x-axis shows ST prediction accuracies, and the y-axis shows MT prediction accuracies: a MT-CV1 using both leaf senescence (LS) and leaf rolling (LR) as secondary traits; b MT-CV2 using both LS and LR; c MT-CV2 using LS only; and d MT-CV2 using LR only. Colors denote seven individual trials and across trials. Values represent means of 100 cross-validation runs Discussion Leveraging genetic diversity is essential for sustainable crop improvement. The DROPS panel (Millet et al. 2016), which has been characterized extensively across a wide range of European environments, represents a valuable resource of dent maize diversity. To further harness the genetic variation observed for our target traits, we selected eight inbred lines with contrasting yield performance under drought and heat stress, ensuring limited variation in flowering time, to develop a MAGIC population. This population was genotyped using whole-genome shallow sequencing and phenotyped across 21 field experiments. We dissected the genetic architecture of important agronomic traits, primarily grain yield, evaluated the usefulness of proxy traits for improving grain yield under stress conditions, and provided recommendations for optimizing MAGIC genotyping and mapping strategies. Evaluating proxy traits for yield improvement The eight MAGIC founders were selected to represent variation in heat and drought tolerance and to capture part of the genetic diversity of the Non-Stiff Stalk heterotic group within the DROPS panel. They exhibited polymorphism for 74% of the markers analyzed in the DROPS panel and were polymorphic for 39 of the 48 grain yield QTL identified (Millet et al. 2016). In our seven testcross trials, the average yield range for the MAGIC population (4.8–10.1 t/ha) covered the broader environmental variation observed across the 29 experiments of the DROPS panel (1.5–11.2 t/ha) only partially. Despite capturing only a subset of the genetic diversity and a narrower environmental range, seven of the 48 DROPS grain yield QTL overlapped with the genomic positions of grain yield QTL detected in this study (Table S6). Evaluating long-established drought proxy traits, such as leaf senescence, leaf rolling, and anthesis-silking interval, offers valuable secondary selection targets (Bänziger et al. 2000). These traits have been incorporated into selection indices for improving yield in drought-prone environments in testcross trials (Edmeades et al. 1999). Here, we measured these traits not only in testcross but also in line per se trials, which can be conducted in short observation plots and facilitate scalable assessment when combined with Unmanned Aerial Vehicle (UAV)-based high-throughput phenotyping (Baret et al. 2018; Cerrudo et al. 2017). This scalability may compensate for the generally low correlation observed between the proxy traits and grain yield. In the testcross trials, we observed modest but significant correlations between grain yield and both leaf senescence and leaf rolling, aligning with Monneveux et al. (2008), who also reported only marginal relevance of conventional secondary traits for yield as a target trait. Despite the low overall correlations, the top 10% highest-yielding MAGIC DH lines exhibited significantly delayed senescence and reduced leaf rolling, suggesting that these traits remain distinguishing features of tolerant lines. The negative correlation between leaf rolling and yield suggests that rolling reflects drought sensitivity rather than a protective mechanism in these genotypes (Parajuli et al. 2018). Finally, the lack of correlation of the anthesis-silking interval with grain yield in testcrosses was likely due to the short interval under stress conditions (2.3 days in MUR21-R), indicating that this trait is already optimized in modern hybrid maize (Araus et al. 2012). Leaf senescence and leaf rolling in line per se trials were correlated with testcross yield only under high-temperature and water-deficit conditions. Although no major pleiotropic QTL were identified, a clear trend emerged among the detected proxy trait QTL: alleles associated with delayed senescence or reduced rolling generally showed positive effects on testcross grain yield, indicating that accumulating these favorable alleles could enhance yield performance. A shared polygenic basis, pleiotropy of minor genes, is further supported by the improved prediction accuracy of grain yield when these two traits were included in the multi-trait GBLUP models. However, given the relatively small gains in prediction accuracy, the additional efforts required for including proxy traits in a breeding context seem justified only if they can be assessed in high throughput and with high precision. QTL for trait improvement In this study, we report several QTL for grain yield, proxy traits, and agronomic traits, which represent valuable targets for trait improvement. Sekhon et al. (2019) reported 64 candidate genes for leaf senescence, via a comprehensive GWAS and gene expression experiment using a diversity panel composed primarily of Stiff Stalk and Non-Stiff Stalk dent lines. We found four of these genes located within our QTL regions (File S2): Zm00001d030968 (Flowering locus K domain protein), Zm00001d052798 (Trihelix transcription factor 3), Zm00001d002967 (AFG1-like ATPase family protein), and Zm00001d020396 (Trehalose-6-phosphate synthase13, trps13). Thus, our results provide a valuable resource for a future study on potential candidate genes in these genomic regions. For leaf rolling, one QTL (qLR(LP)03C) is of particular interest due to its potential pleiotropic effects on yield and its consistent effect across trials. This QTL was detected by all three QTL detection methods. The Lo1056 and B106 alleles reduced leaf rolling (beneficial) with the Lo1056 allele showing the stronger effect (Fig. S8b). This leaf rolling QTL co-localized with a leaf angle QTL identified in the maize NAM population (Tian et al. 2011), suggesting a potential genetic link between the leaf rolling mechanism and overall leaf architecture. While the anthesis-silking interval was relatively short in testcross trials (1.3 days across environments), it averaged 3 days in line per se trials and showed greater variation among genotypes, indicating potential for further improvement. Given the observed negative correlation between anthesis-silking interval and line per se grain yield (rr = − 0.35), the beneficial allele, which shortened the interval, may, therefore, remain important for seed production of inbred lines. Among agronomic trait QTL, a plant height QTL, qPH_final(LP)01A, showed a strong effect (~ 11 cm) across line per se trials but not in testcross trials. The B106 allele significantly reduced plant height. A candidate gene, Brachytic2 (Br2) (Multani et al. 2003), involved in polar auxin transport, lies ~2.2 Mb away from the lead SNP of the QTL. Sequencing data revealed a potential 1 kb duplication within exon 5 of Br2 in the B106 allele, which may contribute to the reduction in plant height. As the Br2 gene has been exploited to develop short-stature maize hybrids, marketed as "smart corn," the B106 allele might also represent a valuable allele for future breeding applications. IBS- and IBD-based methods for QTL detection The known founder genotypes of the MAGIC population enabled genomic mosaic reconstruction and modeling of multi-allelic QTL effects (Li et al. 2021). Odell et al. (2022) recommended applying multiple approaches to obtain comprehensive and complementary results. SNPs have long been used in multi-trait and -environment QTL analysis (Jiang and Zeng 1995; Korte et al. 2012). More recently, this framework has been extended to multi-allelic (FH and AH) approaches. Garin et al. (2020) modeled QTL effects as fixed effects, decomposing them into main effects and Q×E interaction effects. Alternatively, Li et al. (2024) treated QTL effects as random, allowing variances to differ among environments and reduce complexity. Fixed models might lose power as the number of environments or trials increases. The assumption of normally distributed allelic effects in the mixed model might not always reflect biological reality. Here, we applied IBD-based QTL mapping (FH and AH) with a random effect model and compared it with the SNP-based approach. The SNP method identified a greater number of unique QTL in both individual trial and multi-trial analyses. On the other hand, individual QTL explained more phenotypic variance with IBD-based methods, likely due to the estimation of multiple allelic effects; however, the total QTL variance explained per trait was greater with the SNP-based approach (Fig. S10). Given the dense set of 2.7 million SNPs, which likely approaches saturation for capturing all chromosome recombination events, and the broader QTL set identified, our analyses focused primarily on SNP-based results, with IBD-based methods as a complement. Based on these results, we recommend combining SNP-based and IBD-based (FH or AH) methods for MAGIC populations. IBD-based methods are ideal for downstream gene cloning, as they identify fewer but stronger QTL, are less sensitive to individual SNP genotyping errors, and provide founder allelic effects. Furthermore, at low marker densities, they show advantages: When SNPs were reduced to 600 k and 15 k array positions (Method S2), LOD curves from the IBD-based methods remained largely unchanged, whereas SNP detected fewer QTL with lower significance for lead SNPs in the 15 k dataset (Fig. S11). In contrast, at sufficiently high SNP densities, SNP-based QTL detection offers greater power for the detection of QTL with few segregating alleles and is preferable for marker-assisted selection due to the well-established frameworks for marker design, effect estimation, and statistical modeling. Shallow-seq efficiency SNP marker density and precision are determined by the genotyping platform employed. Reduced representation sequencing (RRS), which targets only a small fraction of the genome for sequencing, and fixed-content SNP arrays have been widely used in maize studies due to their high accuracy, low missingness, comparability across experiments, and reasonable costs (Rasheed et al. 2017). These platforms are well suited for genomic prediction and linkage map construction (Chen et al. 2014; Crossa et al. 2013; Ganal et al. 2011). However, both approaches are limited to fixed genomic positions and prone to ascertainment bias, making them less effective for discovering rare variants in diverse germplasm panels (Davey et al. 2011; Romay et al. 2013). Low-coverage whole-genome sequencing (lcWGS), referred to here as shallow-seq, combined with effective imputation, offers more uniform genome-wide coverage and higher marker density (Scheben et al. 2017). With sequencing costs continuing to decline, lcWGS is a practical and scalable option for maize genetics (Lou et al. 2021). The drawbacks of this technology are higher error rates, greater missingness, and limited comparability across experiments, but these disadvantages can be mitigated by optimizing sequencing depth. The optimal sequencing depth can differ depending on population type, founder information, minor allele frequency, and sample heterozygosity (Lou et al. 2021). The rich body of sequencing data generated for the MAGIC (> 50× for the founders and average 5.7× for the DH lines) provided a valuable resource for evaluating and refining genotyping pipelines. Here, we evaluated a scenario with known founder genotypes and examined the number of SNPs detected across sequencing depths. Our results showed that 2× is a turning point, beyond which gains in SNP detection, genotyping rate, and genome coverage begin to plateau. At this depth, 75% of known SNP loci were retained with a missing rate below 0.5, allowing for effective imputation. Given the fixed costs of library preparation, 2× offered the best balance of cost and performance. However, when budgets are limited, maintaining at least 0.5× coverage per sample remains a practical alternative, still yielding over 22 k SNPs, sufficient for IBD-based mapping in this population (Fig. S11). Progressing from the dent diversity panel for crop improvement, we developed a MAGIC population supported by extensive genotypic and phenotypic datasets. This study focused primarily on dissecting the genetic architecture of grain yield and evaluating the utility of drought-related proxy traits for yield improvement. Beyond this, the MAGIC resource allowed to address a broad range of research questions such as optimizing lcWGS and founder detection methods, serving as a training population for marker-based genomic selection, and candidate gene identification for agronomic traits. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (XLSX 192 KB)Supplementary file2 (XLSX 314 KB)Supplementary file3 (PDF 2246 KB) Supplementary file1 (XLSX 192 KB) Supplementary file2 (XLSX 314 KB) Supplementary file3 (PDF 2246 KB)