The incomplete dominance of gene expression controlled by Trans-eQTL hotspots contributes to heterosis in maize Authors: Gen Xu, Xuerong Yang, Mingyue Zhang, Congbin Kang, Ziang Tian, Yunhui Qi, Meijie Luo, Peng Liu, Jeffrey Ross-Ibarra, Jinliang Yang, Hongjun Liu Published: 2026 ABSTRACT AbstractHeterosis, or hybrid vigor, is a key genetic and agricultural phenomenon driven by non-additive effects, though its mechanisms remain debated. To examine these effects at the transcriptomic level, we integrate population-scale transcriptomic and phenotypic data in maize. We find that ∼30% of expressed seedling genes in many hybrids deviate from mid-parent values, consistent with incomplete dominance. While overall expression is largely controlled by cis-eQTLs, non-additive expression is disproportionately regulated by trans-eQTLs. These trans-eQTLs cluster into hotspots that coordinately control hundreds of genes, often within co-expression networks and enriched for transcription factors. We validate one hotspot involving ZmR1, a bHLH transcription factor, showing its role in regulating widespread incomplete dominance and enhancing heterosis in seedling and agronomic traits. These findings support a model in which trans-acting factors shape expression dominance and contribute to heterosis, providing a framework for crop improvement.Subject terms: Plant hybridization, Agricultural genetics, Gene expression profiling Introduction Heterosis, or hybrid vigor, refers to the superior performance of hybrid progeny relative to their parents. Heterosis is widespread in cereal crops1, vegetables2, and livestock3, and has a profound impact on agricultural productivity. The genetic basis of heterosis has been debated for over a century, with three classical quantitative genetic models proposed: dominance (or “pseudo-overdominance”)4–7, overdominance8,9, and epistasis6. Supported by both theoretical and empirical evidence10–12, the dominance hypothesis has recently regained attention, centering on the core idea of directional dominance—that the beneficial effects of dominance tend to act in the same direction13—as well as the complementation of widely distributed and slightly deleterious mutations throughout the genome in subdivided populations14,15. The early theoretical frameworks for the mechanism of dominance laid the critical groundwork for understanding heterosis. Fisher proposed that dominance evolves through selection for “modifier” alleles that suppress the phenotypic effects of deleterious mutations16. Nonetheless, Wright17 mathematically demonstrated that modifiers that increase dominance, or masking of deleterious recessive alleles in hybrids, have only weak selective advantages and therefore are unlikely to be fixed rapidly under natural selection. Instead, Wright argued that dominance arises inherently from nonlinear biochemical interactions in metabolic pathways18. Subsequent studies aligned with Wright’s model, demonstrating that dominance depends on the functional importance of genes and their position in metabolic networks19,20. These insights were later extended by Huber et al.21, who showed that dominance effects correlate with gene expression optimization and dosage sensitivity, suggesting that both evolutionary constraints and molecular mechanisms influence dominance and heterosis. Modern genomic approaches have expanded heterosis research to genome-wide expression studies, with a key focus on distinguishing between additive (parental mid-parent-like)22–26 and non-additive (transgressive or dominant) patterns27–30. However, most studies have been limited by small sample sizes, with the exception of a recent study using a QTL mapping population in maize31. In this study, we investigated the genetic basis of gene expression patterns and their contribution to phenotypic heterosis by generating population-scale RNA-seq data from a large number of hybrids and their inbred parents. Through integrative analyses, we found that many genes exhibit incomplete dominance patterns of expression and identified population genetic signatures associated with this pattern. In particular, we discovered an eQTL hotspot containing ZmR1 locus, a bHLH transcription factor (TF) from the R1 gene family32, which regulates the incomplete dominance of hundreds of genes and is associated with phenotypic heterosis of several traits. Through detailed analysis of overexpression lines, we demonstrated that hybrids overexpressing ZmR1 outperformed their wildtype counterparts. This study provides experimental evidence for understanding heterosis and sheds light on harnessing this classical phenomenon for further crop improvement. Results Incomplete dominance of gene expression in maize seedlings from a blocked diallel population We generated n = 599 hybrids by crossing n = 203 diverse inbred lines according to a blocked partial diallel design (see “Methods”; Fig. 1a, Supplementary Fig. S1 and Supplementary Data S1). The population structure analysis assigned most of these inbred lines to six previously defined heterotic groups33: stiff stalk (SS), non-stiff stalk (NSS), Iodent, group A (PA), group B (PB), and Sipingtou (SPT) (Supplementary Fig. S1a–c). To minimize the confounding effects of genetic dosage and residual endosperm-derived nutrition during early seedling development29, we dissected embryos for germination and seedling trait assessment (Fig. 1b, c and Supplementary Fig. S2a). Consistent with previous studies14,34,35, hybrids outperformed inbred parents across all seven seedling traits (Supplementary Fig. S3), with mid-parent heterosis (MPH) values ranging 12.6%–81.2% for seedling traits (Fig. 1e and Supplementary Data S2), as compared to 10.8%–145.8% for agronomic traits collected from replicated field trials using the same population (see"Methods”; Supplementary Fig. S2b and Supplementary Data S2). Schematic illustration of data collection and heterosis distributions. a Kernels were dissected to remove the endosperm, retaining only the embryo for seedling trait measurements and subsequent RNA-seq experiments. b, c Seedlings germinated from manually dissected embryos were used to collect seedling traits (c). d The two-week-old whole plant tissues were subjected to RNA-seq experiments. e Distribution of mid-parent heterosis (MPH) values for seedling traits. The inset shows the phenotypic distributions of hybrids and their inbred parents for the average dry weight. FW fresh weight, ADW average dry weight, ML mesocotyl length, RL root length, SL seedling length, pSL partial seedling length, TL total length. Phenotypic data were collected from n = 590 hybrids. f Median deviation from the mid-parent value (DMP) for gene expression. Genes are categorized into three groups: positive (yellow), negative (green), and additive (gray). Blue dots represent DMP values from housekeeping genes, and error bars indicate standard deviation. In the upper panel, the genes involved in the benzoxazinoid (Bx) metabolic pathway, with yellow and green colors indicating positive and negative DMP, respectively. Numbers above the violin plots indicate P values from two-sided Wilcoxon tests. Gene expression data were collected from n = 191 inbreds and n = 599 hybrids. Violin plots show the distribution of the data. The white dot indicates the median, the box represents the interquartile range, and the whiskers indicate the minima and maxima. Source data are provided as a Source data file. Using two-week-old whole-seedling tissues, we performed transcriptome profiling of hybrids and their inbred parents (see “Methods”; Fig. 1d and Supplementary Data S3). To ensure comparability across genotypes, RNA-seq libraries were prepared with standardized tissue quantities (or equivalent RNA input). Following RNA-seq data processing (see “Methods”), we identified n = 26, 020 and n = 27, 316 expressed genes (or median FPKM > 1) in the inbred and hybrid populations, respectively. We showed that a set of conserved housekeeping genes36 exhibiting stable expression patterns in both inbreds and hybrids (Fig. 1f and Supplementary Fig. S3), suggesting the feasibility of direct gene-level comparisons between hybrids and their parental lines. Among n = 19, 880 expressed genes, DMP values—defined as the deviation of hybrid expression from the mid-parental expression level—exhibit a broad distribution (Fig. 1f). To detect DMP genes without biological replication, we developed a two-stage likelihood ratio test (2sLRT) method and demonstrated a high concordance with results obtained from replicated datasets (Supplementary Methods and Supplementary Fig. S4 and Supplementary Data S4). Using the 2sLRT method, we found 6458 genes (32.5%) from 599 hybrids significantly deviated from additivity (i.e., show incomplete dominance) (Kolmogorov-Smirnov (KS) test, FDR = 0.05), including 3311 (16.6%) with positive and 3147 (15.8%) with negative DMP values (Supplementary Data S5). For example, genes within the benzoxazinoid (Bx) pathway—a well-characterized and biologically coherent secondary-metabolite pathway37—exhibit variation in the sign and direction of DMP values, likely reflecting their position and function within the pathway (Fig. 1f and Supplementary Fig. S5). Genetic and evolutionary determinants of gene expression patterns Genes exhibiting extreme expression levels in maize inbreds are likely dysregulated due to the accumulation of deleterious mutations38. To investigate this, we quantified the derived deleterious genetic load in inbred lines and found that, indeed, genes with extreme expression levels in the inbred parental lines showed elevated mutational burden in their upstream regulatory regions (see “Methods”, Fig. 2a). Consistent with the role of complementation of deleterious mutations in heterosis, conserved syntenic ortholog genes of maize and sorghum are significantly enriched (permutation test, P values = 0.001) among these genes showing significant DMP values (Supplementary Fig. S6). Genomic features and gene expression networks associated with expression DMP values. a The median genetic load within the 5 kb upstream regions of genes using inbreds. Genes were stratified into 200 equal groups based on their expression levels. The dashed lines define the best-fit regression line, and the gray band represents the 95% confidence interval. b Gene expression DMP values for hub and non-hub genes in selected modules. The x axis indicates module IDs. P values from two-sided Wilcoxon tests are shown above the violin plots, and gene numbers are shown below. The white dot indicates the median, the box represents the interquartile range, and the whiskers indicate the minima and maxima. c A subset of the co-expression network of M19. Nodes (dots) vary in size and color, representing different gene categories. d Lignin and anthocyanin pathways. Genes marked with different colors represent distinct gene categories. Source data are provided as a Source data file. We constructed coexpression networks using hybrid gene expression data and identified 19 distinct gene modules (Supplementary Fig. S7). Within each module, the hub genes show more pronounced non-additivity compared to the non-hub genes (Fig. 2b). The largest module (Module 19, comprising n = 1, 813 co-expressed genes) predominantly consists of genes exhibiting positive DMP values and is enriched for genes involved in the hydrogen peroxide catabolic process or genes in the lignin metabolic pathway (Supplementary Fig. S7, see results for other modules in Supplementary Data S6). Module 19 contains a number of anthocyanin genes (n = 17, Supplementary Data S7), which are a part of the lignin pathway (Fig. 2c, d), and genes in the benzoxazinoid (n = 9) pathway (Fig. 1c), as well as homologous genes previously reported to contribute to heterosis in rice (Fig. 2c), such as TAC1 and GHD839. The trans-eQTL disproportionately regulates incomplete dominance of gene expression To assess the genetic basis for the incomplete dominance of gene expression, we used DMP values to perform an eQTL analysis40,41. As a result, we identified n = 14,477 eQTLs associated with DMP values, as well as n = 19,079 and n = 12,942 eQTLs for gene expression per se in hybrid and inbred populations, respectively (see “Methods”; Supplementary Data S8). Following previous studies26,38,42, we defined eQTLs located within a 1-Mb flanking region of the target gene as cis-eQTLs. Those located more than 1 Mb away on the same chromosome were classified as trans1, and those located on different chromosomes were classified as trans2, to distinguish potential long-distance linkage or regulatory effects on the same chromosome from unlinked regulatory loci. For example, the Starch-Branching Enzyme (SBEIIa), a gene involved in the amylopectin biosynthesis pathway43, has a cis-eQTL along with eight trans1- and five trans2-eQTLs detected in the hybrid population (Fig. 3a). We detected seven trans1- and 28 trans2-eQTLs associated with the DMP value of SBEIIa expression, including several TFs likely involved in the regulation of starch biosynthesis (Supplementary Data S9). The cis- and trans-eQTLs for gene expression DMP values. a Manhattan plots of GWAS for gene SBEIIa expression level in inbred, hybrid populations and the DMP values. Genes with blue color are transcription factors. The horizontal dashed line denotes the genome-wide significance threshold (2.9 × 10−6 for the inbred population and 5.4 × 10−6 for the hybrid population), determined based on the effective number of independent SNPs (Neff; significance threshold = 1/Neff) to correct for multiple testing. b Proportions of eQTLs located in each genomic feature. Trans1 refers to trans-eQTLs located on the same chromosome as their associated genes, while Trans2 refers to those located on different chromosomes. Up5k, within 5 kb upstream of the target gene; Down5k, within 5 kb downstream of the target gene; Up5k–1 M, region spanning from 5 kb to 1 Mb upstream of the target gene; Down5k–1 M, region spanning from 5 kb to 1 Mb downstream of the target gene. c Comparison between eQTLs and histone modification peaks in the B73 genotype. The red asterisks indicate that eQTLs are significantly enriched in the histone modification peaks. P values from one-sided permutation tests are shown above the violin plots. d Dominance effects (d/a value) of eQTLs for gene expression traits per se in hybrids (upper panel) and for the transformed DMP values (lower panel). Orange and green lines indicate results from cis and trans-eQTLs. Source data are provided as a Source data file. A further summary of the eQTL analysis revealed that cis-eQTLs associated with expression per se were identified for over 70% of genes in both inbred lines (79.7%) and hybrids (77.3%), with most located within 5 kb upstream or downstream of the corresponding gene. In contrast, only 2.4% of genes with significant eQTL for DMP values exhibited associations in cis, whereas 89.5% were located on chromosomes different from those of their associated genes, suggesting that the incomplete dominance expression pattern is more likely to be regulated in trans2 (Fig. 3b). These results are insensitive to our definitions of the cis–trans boundary (see results across thresholds ranging from 200 kb to 2 Mb in Supplementary Fig. S8). Both cis- and trans-eQTLs are significantly enriched in genomic regions marked by the active histone modifications H3K4me4 and H3K27ac (Fig. 3c), suggesting that genomic variants reside in open chromatin regions with the potential to influence transcription. To evaluate the genetic effects of eQTLs on expression, we further estimated the dominance index (the ratio of dominance to additive effects, d/a) for each eQTL detected in the hybrid population (Fig. 3d). As expected for expression levels per se, both cis- and trans-eQTLs (including both trans1 and trans2) showed dominance index values centered around 0, consistent with predominantly additive effects. In contrast, for DMP values, while cis-eQTLs remained centered near 0, trans-eQTLs displayed a bimodal distribution, with peaks indicative of incomplete recessive and incomplete dominance, supporting that trans-regulation drives deviation from additive expectation in gene expression. Most of these deviations represent partial or incomplete dominance rather than strict complete dominance, consistent with the quantitative variation in the effects of gene dosage21. The trans-eQTLs cluster at specific genomic hotspots In summarizing eQTLs, we found that eQTLs for gene expression per se are widely distributed throughout the genome, while eQTLs for DMP values are concentrated in specific hotspots (Supplementary Fig. S9). We identified n = 822 eQTL hotspots, each regulating the incomplete dominance of more than 45 genes (Fig. 4a, Supplementary S10, and Supplementary Data S10). Of these, 675 (82%) eQTL hotspots are enriched for target genes with positive DMP values, while only 92 (11%) are enriched for genes with negative DMP values. We found that eQTL hotspots were significantly less likely to overlap with genomic regions under positive selection during maize breeding between the 1960s and the 1980s44 than expected by chance (permutation test, P = 0.01), suggesting that the maintenance of genetic diversity at these hotspots may be important for heterosis. eQTL hotspots for gene expression DMP values. a Distribution of the target gene number of eQTL hotspots. The red triangles represent the physical position of three candidate genes (Zm00001d039267, Zm00001d048623, and Zm00001d025281) under eQTL hotspots. Yellow and green dots represent target genes of eQTL hotspots enriched for positive and negative DMP values, respectively. b DMP values of the target genes of the three highlighted genes across the genome. c–e Examples show three hotspots where target genes overlap with DAP-seq and ChIP-seq peaks. The upper panels show the GWAS signals around the target genes for expression per se and expression DMP. The horizontal dashed line denotes the genome-wide significance threshold (2.9 × 10−6 for the inbred population and 5.4 × 10−6 for the hybrid population), adjusted for multiple testing. The red and gray boxes indicate the target and other gene models. The middle panels show the target gene structure. The blue arrows indicate gene orientation. The bottom panels show the DAP-seq (c, e) and ChIP-seq (d) signals (green color) of the trans-eQTL candidate TF for each target gene. Source data are provided as a Source data file. Certain genes located within trans-eQTL hotspots, including those encoding TFs, may act as key regulators of transcriptional activity. To test this hypothesis, we retrieved maize TFs from the PlantTFDB database45. Enrichment analysis revealed that TFs are significantly overrepresented within trans-eQTL hotspots (permutation test, P = 0.004). Two TFs, Zm00001d039267 (Auxin response factor 7) and Zm00001d048623 (MYB-transcription factor 88), underlying two eQTL hotspots on chromosomes 3 and 4 (Fig. 4a), have public DAP-seq46 and ChIP-seq47 data, separately. We processed the DAP/ChIP-seq data and compared the target genes of the two hotspots with the binding sites of the TFs. The analysis revealed that TF binding sites are significantly (permutation test, P = 0.001) overrepresented among the target genes of trans-eQTL hotspots—by 60% and 40% more than expected by chance for Zm00001d039267 and Zm00001d048623, respectively (Fig. 4b–d). Furthermore, we generated DAP-seq data for another TF gene, Zm00001d025281, annotated as AP2/EREBP, underlying a trans-eQTL associated with DMP values for 96 genes, 67 (69.8%) of which significantly overlapped with DAP-seq peaks (permutation test, P = 0.001; Fig. 4b, e). Gene expression contributes to seedling trait heterosis To further assess the impact of gene expression on phenotypic heterosis, we conducted transcriptome-wide association studies (TWAS) for seedling trait heterosis (see “Methods”). In total, 54 genes were associated with seedling traits per se and 25 for seedling heterosis, with n = 11 genes (44%) overlapping between the two sets (Fig. 5a). A large proportion of these trait-associated genes, particularly those associated with heterosis traits, are TFs. For example, the MADS-box gene TU1, previously implicated in the regulation of grain yield and leaf number above the ear48, was associated with both seedling fresh weight and its heterosis (Fig. 5a, see other genes and their potential functions in Supplementary Data S11). Another TF ZmR1, a bHLH that participates in the synthesis of anthocyanin pigmentation49,50, was strongly associated with one seedling trait per se (FW) and four seedling heterosis traits (FW, pSL, SL, and TL) (Fig. 5a). The expression level of ZmR1 is higher in most hybrids compared to their parental lines (Fig. 5b), exhibiting a significant deviation from the mid-parent value (paired Student’s t test, P = 2.9e-49; Supplementary Fig. S11). Furthermore, this gene is located within a trans-eQTL hotspot and potentially regulates the DMP values of more than 230 genes, the majority of which show positive DMP values (Fig. 5c). We then investigated whether these targets are correlated with phenotypic performance. Indeed, correlation analyses revealed that expression levels of most ZmR1 target genes were positively correlated with the four seedling traits, and their DMP values were likewise positively correlated with heterosis in these traits (Supplementary Fig. S12). Notably, several of these targets are previously characterized genes, including FHT1, involved in the anthocyanin biosynthesis pathway51; CESA2, annotated as a cellulose synthase52; and two photosystem-related genes — CAB353 and LHCB154. Consistent with the regulatory role of the eQTL hotspot, genes co-expressed with this candidate gene exhibit even stronger non-additive expression pattern (Fig. 5c). Co-expression network analysis further reveals that the target genes of this eQTL hotspot are distributed across multiple modules, particularly in module 19, where they are co-expressed with a set of genes involved in lignin, anthocyanin, and benzoxazinoid biosynthetic pathways (Fig. 5d). To validate the downstream targets of ZmR1, we randomly selected five potential target genes for a dual luciferase assay (see “Methods”). The results demonstrated that ZmR1 functions as a transcriptional activator for Zm00001d021599, Zm00001d013133, and LHCB3, while it exerts transcriptional repression on Zm00001d042730 and Zm00001d005478 (Fig. 5e). ZmR1, a candidate gene within the trans-eQTL hotspot, is associated with seedling trait heterosis. a Transcriptome-wide association studies (TWAS) for seedling traits per se and their heterosis. Gene names highlighted in green are associated with both traits per se and heterosis. Dashed horizontal lines indicate the Bonferroni-corrected significance threshold for TWAS (P = 1.8 × 10−6), based on 0.05 divided by the number of tested genes. b Expression level of ZmR1 in hybrids (n = 582) and their parental lines. c DMP values of gene expression regulated by or co-expressed with ZmR1 in inbreds and hybrids. Numbers below the violin plots indicate gene numbers. Two-sided Wilcoxon tests were performed to compare ZmR1 target genes and ZmR1 co-expressed genes with background genes (all genes). d Co-expression network constructed from genes co-expressed with the targets of ZmR1. Node size and color indicate different gene categories. e Validation of ZmR1–target gene interactions using a dual-luciferase assay. For each target, the left image shows the empty vector control; the right shows co-expression with ZmR1. Each violin plot represents n = 6 biologically independent transient expression assays. In (c, e), P values from two-sided Wilcoxon tests are shown above the violin plots. The white dot indicates the median, the box represents the interquartile range, and the whiskers indicate the minima and maxima. Source data are provided as a Source data file. Overexpression lines of ZmR1 influence trait heterosis To quantify the phenotypic effects of ZmR1, we obtained three overexpression (OE) lines in the Jing724 background50. We generated hybrids by crossing an elite inbred line, Jing92, with the wildtype Jing724 and each of three OE lines. Both the OE lines and their hybrids, as previously reported50, displayed a distinct purple coloration in the seedling leaves (Fig. 6a). To evaluate phenotypic outcomes, we measured four seedling traits in a replicated greenhouse experiment (see “Methods”). Overall, the OE allele in Jing724 background exhibited similar or reduced trait values in inbred lines but increased trait values in hybrids (Supplementary Fig. S13a). Furthermore, OE hybrids exhibited progressively enhanced heterosis for fresh weight, seedling length, and root length (Fig. 6b), corresponding to increasing levels of ZmR1 expression (Supplementary Fig. S14). Results from replicated field experiments indicated that OE hybrids exhibited significant increases in the trait per se (Supplementary Fig. S13b), along with enhanced heterosis for plant height and stalk stiffness (Fig. 6b and Supplementary Fig. S15). ZmR1 affects seedling and adult trait heterosis. a The overexpression (OE) lines of ZmR1 and their hybrids. Scale bar = 5 cm. b Seedling and adult trait heterosis in hybrids derived from crosses with OE lines. Adult traits were collected from two locations: Hainan and Tai’an. FW fresh weight, SL seedling length, ML mesocotyl length, RL root length, PH plant height, MSS middle stalk stiffness. The full names of inbred lines J724 and J92 are Jing724 and Jing92, respectively. Comparisons between the OE lines and their corresponding wildtypes were performed using two-sided Wilcoxon tests. P values are shown above the violin plots, and sample sizes are shown below. The white dot indicates the median, the box represents the interquartile range, and the whiskers indicate the minima and maxima. Source data are provided as a Source data file. Since stalk stiffness is a critical trait associated with lodging resistance55, and lignin is a major contributor to cell wall rigidity56, we further examined lignin accumulation in stem tissues. We performed phloroglucinol staining (see “Methods”), which specifically detects lignin57, and found that OE lines, particularly OE hybrids, exhibited more intense staining, indicating elevated lignin content (Fig. 7a). These visual observations were consistent with quantitative measurements of lignin content, which showed significantly higher levels in the OE hybrids relative to their matched wildtypes (Fig. 7b). Furthermore, MPH for lignin content was significantly increased in hybrids derived from OE lines (Fig. 7c). Overexpression of ZmR1 enhances lignin accumulation and expression non-additivity. a Phloroglucinol staining of stem cross-sections from ZmR1 OE lines. Scale bar = 500 µm. Representative images from three independent biological replicates with similar results are shown. b, c Lignin content (b) and its heterosis (c) in ZmR1 OE lines. Comparisons between the OE lines and their corresponding wildtypes were performed using two-sided Wilcoxon tests. P values are shown above the violin plots, and sample sizes are shown below. The white dot indicates the median, the box represents the interquartile range, and the whiskers indicate the minima and maxima. d Comparison of expression DMP values between hybrids generated from ZmR1 OE lines and the corresponding wildtype hybrids. Dark red dots represent ZmR1 target genes showing a non-additive pattern, and orange dots indicate genes involved in the same pathway as ZmR1. e Log2-transformed expression fold changes (log2FC) between OE lines and their matched wildtypes for ZmR1 eQTL target genes and genes involved in lignin and anthocyanin biosynthetic pathways. f A proposed model illustrating the elevated non-additivity of gene expression. A trans-eQTL can regulate the expression of its target gene and downstream genes in the hybrid. “*“ represents the alternative allele introduced by mutation. Source data are provided as a Source data file. At the transcriptome level, most target genes of ZmR1 associated with the eQTL hotspot exhibited increased DMP values in OE hybrids (Fig. 7d). We further assessed the expression patterns of genes co-expressed with ZmR1 and found a similar pattern, with most genes showing increased DMP values in OE lines, both in inbreds (Supplementary Fig. S16a) and hybrids (Supplementary Fig. S16b). In addition to ZmR1, several key genes in the lignin and anthocyanin biosynthesis pathways, such as HCT7, PAL5, A1, A2, BZ1, and AAT1, also showed elevated expression per se (Fig. 7e) and increased DMP values (Fig. 7d) in OE hybrids. Taken together, our results suggest that ZmR1 shifts incomplete-dominance gene expression in a consistent direction (i.e., increasing DMP values) for hundreds of genes at the transcriptomic level, leading to phenotypic heterosis in both seedling and adult traits, particularly for lignin-related traits. Discussion In this study, we used MPH as a proxy for trait heterosis, as it can be modeled as the sum of dominance effects (d) weighted by squared allele frequency differences (y2) between two parental lines across loci genome-wide58. In this framework, y represents the difference in allele frequency between populations, that is, two different heterotic groups, Stiff Stalk and Non-Stiff Stalk59, and y2 equals one only when parents are fixed for alternate alleles. This model highlights that directional dominance, in which dominance effects consistently favor the same direction, is required for positive heterosis. Consequently, the amount of heterosis is specific to each individual cross for a given trait13, depending on the pattern of allele divergence and dominance interactions throughout the genome, consistent with our observations for the heterosis of both seedling and adult traits, as well as previously published results14,60. Analogous to MPH, we applied DMP values to gene expression rather than organismal traits. Our results indicated that approximately 30% of expressed seedling genes (n = 6458) exhibited DMP values that significantly deviated from zero, primarily reflecting incomplete dominance, in >5.1% of hybrids, while only about 500 genes exhibited this pattern across the majority of hybrids (>50%). This suggests that most genes across the majority of hybrid families follow an additive expression pattern, although a substantial proportion exhibit incomplete dominance in some hybrids31. However, further validation is warranted due to technical limitations, including the lack of replication in our population-scale RNA-seq data and the limited set of housekeeping genes used in our modeling. If we further hypothesize that phenotype is a function of genome-wide gene expression variation under an omnigenic inheritance model61, phenotypic heterosis should reflect the cumulative effects of expression-level dominance (or incomplete dominance) acting in the same direction. This idea aligns with the observation that co-expressed genes, possibly involved in the same metabolic pathways (e.g., the Bx or lignin pathways), tend to exhibit coordinated patterns of positive or negative dominance. The cis-regulation alone, while accounting for the majority of detectable eQTLs for the trait per se in hybrids, rarely causes gene expression to deviate from additive inheritance patterns, and its ability to evolve over generations appears limited. In fact, despite decades of improvement in yield, the magnitude of heterosis in maize has remained largely constant62,63. Consistent with this pattern, we found that eQTL hotspots show little evidence of selection during modern breeding, suggesting that these regions may be maintained to preserve regulatory diversity underlying heterosis. Non-additive gene expression is therefore unlikely to arise from variants within the gene or its immediate upstream region. This is consistent with our finding that disproportionately more trans-eQTLs were detected for DMP values, and that non-additive expression is largely controlled by heterozygous regions of the genome31. At the molecular level, trans factors may function as diffusible proteins that bind to functional alleles in cis, recruiting the complete gene transcription machinery64, thus masking the deleterious effects of cis-regulatory variants (Fig. 7f)—a model in line with the dosage effects of TFs65. Consistent with the omnigenic model61, the expression levels of trans-eQTLs (peripheral genes) do not necessarily have direct effects on phenotypic heterosis (or are difficult to detect due to the relatively small effect size). Instead, they can indirectly affect heterosis by modifying the expression patterns of hundreds of core genes. Such trans-eQTL modification in specific genes would have limited fitness effects and likely would not evolve, but modifiers that affect hundreds or thousands of genes could confer a selective advantage66. In support of this idea, a key insight from our study is that incomplete dominance of gene expression in hybrids is primarily governed by trans-acting regulatory loci, i.e., the trans-eQTL hotspots. In particular, many of these hotspots contain TF genes, suggesting that they act as hub genes or master regulators orchestrating coordinated expression deviations from additivity, predominantly in the same direction. These observations suggest an oligogenic regulatory architecture at the upstream level, where a limited number of trans-acting loci influence large downstream gene networks. We perform functional validation of the candidate TF ZmR1 and its downstream target genes using molecular, reverse genetic, and transcriptomic analyses. The ZmR1 gene is located within a major trans-eQTL hotspot and regulates the DMP values of over 230 downstream genes, many of which are part of co-expression networks, exhibit mostly positive incomplete dominance, and are involved in anthocyanin, lignin, and photosynthesis pathways. Functional validation through dual-luciferase assays and expression analysis in overexpression lines confirmed that ZmR1 acts as a transcriptional activator or repressor for specific targets, and that its activity enhances the non-additive expression patterns of these target genes. Overexpression of ZmR1 showed negative effects in inbred backgrounds but positive effects in hybrids, suggesting that regulatory balance is strongly influenced by the heterozygosity of the genetic background. It is important to note that ZmR1 should not be regarded as a single “major heterotic gene”; rather, it serves as a representative example of a trans-acting regulatory factor that influences the expression patterns of hundreds of downstream genes through their cis-regulatory variation (i.e., when these genes are in a heterozygous state that allows the regulatory effect to manifest). Collectively, our results provide empirical evidence that, at the gene expression level, regulatory factors—particularly TFs within trans-eQTL hotspots — may buffer deleterious alleles in the cis-regulatory regions of their target genes. A master TF typically regulates a set of co-expressed genes in a manner of directional dominance, which can ultimately manifest as phenotypic heterosis in a genotype-specific fashion. It is important to note that heterosis in this context can be confounded with natural variation in developmental traits. In our hierarchical model, the regulatory network governed by the master regulator drives trait heterosis only when peripheral genes exhibit regulatory plasticity that allows their expression to deviate from strict additivity in hybrids. This proposed idea aligns with theoretical models of dominance rooted in biochemical nonlinearity and gene network buffering. Taken together, our study provides a predictive framework for identifying and engineering regulatory elements to enhance heterosis in breeding programs. Methods Plant materials for blocked partial diallel mating design and field trial We selected a diverse set of 203 maize inbred lines to generate hybrid crosses. The selected inbred lines represent a subset of the diverse lines of maize HapMap3 and are also part of the panel of global diversity of maize59,67. We grouped the inbred lines into 10 blocks based on their flowering time, and within each block, we made crosses following a partial diallel mating design, resulting in about 600 hybrids. Using these hybrids and their inbred parents, we conducted replicated field trial in multiple locations in China during the summer of 2021: Tai’an, Shandong (36°11’07“N, 117°08’05“E); Suihua, Heilongjiang (SH, 46°39’32“N, 126°58’05“E); and Harbin, Heilongjiang (HRB, 45°51’38“N, 126°49’03“E). Field phenotypic data collection We measured plant height and ear height from at least six plants per line after pollination. After harvest, we collected grain weight, moisture level, and hundred kernel weight using a seed counter. Grain weight was adjusted to 14% moisture, and average grain yield was calculated by dividing the total grain weight by the number of plants in that line. Procedure for embryo dissection We sterilized the seeds with 75% v/v ethyl alcohol for 1 min, rinsed three times with sterile water, then with 10% NaClO for 10 min, and rinsed with sterile water three times. The seeds were then stored in sterile seed bags at 4 °C for further treatment. Subsequently, the sterile seeds were soaked in sterile water overnight for 18 h and then used to dissect embryos from other components of the seed, including endosperm and seed coat. The dissected embryos were then sterilized with NaClO 7% v/v and rinsed three times with sterile water. Every 12 embryos of each hybrid line were covered in sterile brown paper and hydroponically cultured in Hoagland’s nutrient solution (900 ml) in each container, following a previous study68, and 24 embryos (every 12 embryos in 900 ml of nutrient solution) for each inbred line due to the lower germination rate. The embryos were cultured in a growth chamber at 50% relative humidity in an 18 h light/6 h dark photoperiod. To ensure consistent growth conditions, sterile water was added to up to 900 ml of each container daily for 14 days. Then, at 8:00 a.m. on day 14, the individual plant was rinsed with sterile water and dried with sterile filter paper. Finally, lines with at least four growth-consistent individuals were chosen for data collection. Three individual plants from each line were snap-frozen with liquid nitrogen and then stored at −80 °C for RNA extraction. Otherwise, the remaining individual plants were used to measure root and seedling traits. RNA preparation and RNA-seq library construction Total RNA was extracted and pooled from three individual plants for each line using TRIzol ReagentTM (Life Technologies Corporation, Carlsbad, CA, cat. #15596026) according to the manufacturer’s protocol. Then, total RNA was qualified and quantified with a cutoff of 28S/18S > 1.0 and RNA Integrity Number (RIN) ≥ 6.0 by using Agilent 2100 Bioanalyzer (Agilent, CA, USA). Subsequently, the qualified and quantified RNA was used for RNA-seq library construction. Libraries were prepared according to Illumina Standard total RNA-seq library preparation kit. Briefly, total RNA was purified and fragmented, then was used to synthesize first strand and second strand cDNA; after purification, the second strand cDNA was ligated with adapters. After further purification, the DNA fragments with 250−300 bp were enriched and purified as RNA-seq libraries. Finally, libraries were sequenced to generate 150-bp paired-end reads (PE150) using the DNBseq platform from BGI-Shenzhen. Gene expression quantification of RNA-seq data Reads were first filtered using fastp 0.20.069 with default parameters to remove low-quality reads. All filtered reads were then mapped to the maize B73 AGP v4 reference genome70 using the aligner STAR 2.7.9a71 with default parameters. The bam files were then counted from raw reads assigned to reference gene models using featureCounts v2.0.272. Raw read counts were then normalized using the TMM (trimmed mean of M values) normalization method implemented in edgeR73 to give CPM (counts per million) values for each gene. CPM values were then normalized by gene length to give FPKM (fragments per kilobase of exon per million reads) values. Genes with FPKM values < 0.1 in all samples were excluded, and those with FPKM > 1 were regarded as expressed, as described previously26. Calculation of mid-parent heterosis (MPH) for phenotypic traits and deviation from the mid-parent value (DMP) for gene expression The middle-parent heterosis (MPH) for each phenotypic trait was calculated using the formula MPH = (F1 − mp)/mp, where F1 represents the phenotypic value of the hybrid and mp is the average value derived from the two parental lines. Similarly, for gene expression, the deviation from the mid-parent value (DMP) was calculated using the same formula as MPH. Only genes with FPKM values greater than 1 in both parental lines and their corresponding hybrids were included in the analysis. Additionally, each gene’s DMP value was retained only if the missing rate across the entire population was below 30%. We chose to quantify gene expression dominance using mid-parent deviation (DMP) rather than high-parent deviation (DHP) because DMP provides a statistically balanced measure of deviation from additive inheritance. To identify genes that exhibit DMP for each hybrid, we developed a two-stage likelihood ratio test (2sLRT) for unreplicated large-scale RNA-seq count data and applied the test to obtain a P value for each gene in each hybrid. If a gene is additive across all families, we expect the P values for this gene to follow a Uniform (0,1) distribution across families. A one-sided KS test was applied to identify genes whose distribution of P values was stochastically smaller than Uniform (0,1), corresponding to genes that tend to be non-additive across families. The Benjamini-Hochberg method was applied to P values from the KS test to control the false discovery rate (FDR) at 0.05. The proportion of all hybrids for a gene to be detected non-additive at an FDR of 0.05 was also calculated, and this proportion is at most 5.18% among genes with an insignificant KS test. We declare a gene tends to be non-additive based on both FDR control of the KS test and the threshold of minimal proportion of non-additive families (5.18%). For more details, see (Supplementary Methods). A gene is classified as a positive non-additive gene if the median value of its expression of DMP in the hybrid population is greater than zero; otherwise, it is considered a negative non-additive gene. Gene expression QTL (eQTL) analysis Although biological replication was not feasible at this population scale, the large number of genetically diverse lines enabled robust population-level inference for eQTL mapping and expression–phenotype associations. Whole-genome sequencing (WGS) SNPs of parental lines were obtained from the Hapmap 3.2.1 reference panel67, which can be downloaded from https://www.panzea.org/genotypes. Then, the hybrid genotypes were inferred by CreateHybridGenotypesPlugin in TASSEL v5.074. After filtering out SNPs with minor allele frequency (MAF) <0.05, approximately 21 and 11 million SNPs were retained in the inbred and hybrid populations, respectively. We performed eQTL mapping for each gene in the inbred and hybrid populations, as well as expression DMP, using a mixed linear model implemented in GCTA software75. To account for inferred confounders that influence expression variation, we applied the R package PEER76 to estimate the hidden factors using the gene expression matrix from inbred and hybrid populations, respectively. In the analysis, the hidden factors were used as covariates together with the first three principal components (PCs) calculated from genome-wide SNPs using PLINK 1.9 software77. The kinship matrices were calculated using the centered IBS method implemented in TASSEL 5.274. The log2-transformed FPKM values were used as phenotypes in inbred and hybrid populations. The threshold for the significant association SNPs was set to 2.9 × 10−6 (1/n, n = 339, 202 is the number of independent SNPs) for the inbred population and 5.4 × 10−6 (1/n, n = 183, 864) for the hybrid population and expression DMP. Here, the independent SNP number was determined by using PLINK 1.9 with the indep-pairwise option, as described in previous studies78,79 (window size 100, step size 10, r2 ≥ 0.1). We defined eQTL peaks as those identified by considering a 50 kb window upstream and downstream of significant SNPs. Overlapping regions were merged, and the regions with more than three significant SNPs were defined as high-confidence association loci. We defined an eQTL as trans when the physical distance between the mapped QTL and the target gene exceeded 1 Mb; otherwise, the eQTL was classified as cis. Although linkage disequilibrium (LD) in maize decays rapidly80, structural variation and distal promoter or enhancer elements can extend regulatory influence over several hundred kilobases. Therefore, a 1 Mb window was used as a conservative upper bound to minimize false cis assignments. To define an eQTL hotspot, we narrowed each eQTL to 20 kb size according to up and downstream 10 kb of a leading SNP within an eQTL. We merged the narrowed eQTL if detected from different genes and overlapped. Finally, if an eQTL is associated with gene expression DMP in more than 44 genes (top 5%), we define this eQTL as a hotspot. We estimated the dominance index (the ratio of dominance to additive effects, d/a) for each eQTL identified in the hybrid population, following the method described by ref. 26. For each eQTL, the median gene expression or DMP value associated with each allele of the lead SNP was calculated. Dominance (d) and additive (a) effects were then derived using the formulas: d = F1 − mp and a = HP − mp, while F1 is the phenotypic value of the hybrid, mp is the mid-parent value, and HP is the value of the high parent. The level of dominance effect for each eQTL was determined as in our previous study41: overdominance, d/a > 1.25; dominance, 0.25