Using the single-seed descent method, the RIL population, comprising 204 F₈ lines, was developed from a cross between DH1M and T877. In 2017, the experiment was carried out in a completely randomized block design with three replicates at Sanya (18°23′N, 109°44′E) in China. Each line had 78 plants arranged in plots of six rows, with each row being 3.0 meters long and spaced 0.5 meters apart, containing 13 plants per row. Pollination dates for individual RILs were recorded to schedule sampling times. In previous study [15], seed dry weight were recorded at 14 time points post-pollination (10, 15, 20, 25, 30, 35, 40, 43, 46, 49, 52, 55, 58, and 61 days after pollination), covering maize filling stage. We found that 30 DAP marked a period of rapid grain filling, while the period from 40 to 61 DAP exhibited significant variation in grain weight. Therefore, after 30 DAP, seed were taken approximately every 10 days (30, 40, 52 and 61 DAP) to measure content of protein, oil and starch [15]. For each sampling time, two similar ears for each line were selected, and the seeds were dried at 105°C for 1 h, then at 70°C–80°C to constant weight. Approximately 200 randomly chosen seeds were analyzed by a Infratec^TM 1241 Grain Analyzer (FOSS, Hillerød, Denmark) for oil, protein and starch content [9]. The calibrations used in this study were developed by the manufacturer for maize seed, having a determination coefficient of 0.97 for protein, 0.97 for oil and 0.79 for starch based on 1661, 1436 and 955 samples, respectively. The descriptive statistics and phenotypic correlation of GPC, GOC, and GSC at four measurement times were implemented using the R software.

The RIL populations were genotyped using the MaizeSNP50 chip containing 56,110 SNPs. To remove redundant markers, a sliding-window approach [16] was employed to construct a bin map. The order of bin markers was verified with the ripple function in the qtl package [17]. Genetic distances between bin markers were calculated using the “Kosambi” function [18].

The observed GPC, GOC, and GSC at four different times were utilized for QTL analysis. QTL mapping was performed with a windows size of 10 cm and a step size of 1 cm by composite interval mapping (CIM) in R/qtl (v1.50) [17]. An empirical LOD threshold of 3.0 was set, and the support interval for each QTL was defined by a 1-LOD drop on either side of the peak. QTL were named following this format: q + trait name + chromosome + QTL number [19].

To pinpoint candidate gene, genetic and physical maps of QTL were integrated. All the potential candidate genes in QTL interval were identified. The reference genome based on the maize B73 RefGen_V3 genome. Gene functions were assigned using annotation information from the maizeGDB database (http://www.maizegdb.org, accessed on 10/05/2024).

To investigate the expression profiles of candidate genes within QTL, we systematically collected seed samples from T877 and DH1M lines at developmental stages of 10, 20, 30, 40, and 50 DAP. Seeds from the central portion of two comparable ears were selected for RNA isolation and subsequent RNA-seq analysis [15]. RNA-seq library was constructed using a TruSeq RNA sample preparation kit (Illumina). By removing adapters, low-quality reads, and reads containing poly-N from the raw data, the clean data were then aligned to the sequence of B73 RefGen_V3 reference genome using HISAT2 (V2.1.0). Read counts mapped to each gene were determined using FeatureCounts in the Subread package (v1.6.5). The FPKM (fragments per kilobase of exon model per million mapped reads) of each gene was calculated. Differentially expressed genes (DEGs) were identified using the R package ‘edgeR’, with criteria set at an adjusted p value < 0.05 and |log₂(fold change)| ≥ 1.

Whole genome sequencing of two parental line were conducted using Illumina HiSeq 2000 [20]. Clean reads were aligned to the B73 reference genome (ftp://ftp.ensemblgenomes.org/pub/plants/release-24/fasta/zea_mays/dna/, accessed on 10/05/2024) using the BWA-MEM aligner (v0.7.17), which provides accurate alignments for both short and long reads. SNP calling was performed by SAMTools (v1.14). After filtering out low-quality SNPs, SNPs were annotated using snpEff (v4_5) and 19,155 nonsynonymous SNPs in exons were identified between T877 and DH1M. Differentially expressed genes in QTL interval with nonsynonymous variants were selected as candidate genes.

Three grain compositions, including protein (GPC), oil (GOC) and starch content (GSC), were determined by near-infrared (NIR) spectroscopy in the RIL population during the grain filling stage. Significant differences between the two parents were observed at 30, 40, 52 and 61 DAP: DH1M showed higher GPC and GOC, but lower GSC than T877 (Table 1). Within the RIL population, considerable phenotypic variations of the three grain quality traits were observed, with the coefficient of variation (CV) ranging from 5.84% to 6.73%, 8.37% to 12.30% and 1.63% to 2.24% for GPC, GOC and GSC, respectively. During the grain filling stage, GPC and GSC generally increased, while GOC decreased, with a significant change observed from 30 to 40 DAP (Fig. 1). A significant positive correlation was observed between different time points for the same trait. Additionally, protein and oil content exhibited a significant positive correlation, while starch content showed a significant negative correlation with both protein and oil content (Fig. 2).

Between the two parental lines, 9780 high-quality polymorphic SNPs were allocated to 3227 bin markers. These markers were used to construct a genetic linkage map of the DH1M × T877 RIL population (Fig. S1). The genetic map spanned a total length of 2450 cm with an average marker interval of 0.76 cm. The number of markers per chromosome ranged from 111 on chromosome 2 to 503 on chromosome 1. The sizes of the linkage groups varied, with chromosome 2 being 102.29 cm and chromosome 1 extending to 373.06 cm (Table S1). This genetic map provides a foundation for mapping QTLs associated with grain quality traits in the RIL population.

Across all 10 maize chromosomes, a total of 32 QTLs related to the traits studied during the grain filling stage were identified, including 14, 12 and 6 QTLs for GPC, GOC and GSC, respectively (Table 2 and Fig. 3). These QTLs could explain 1.59% to 19.39% of the phenotypic variation, with 8 QTLs having a phenotypic contribution exceeding 10%. The number of QTLs identified on each chromosome varied between one and five. Chromosome 1 had the most QTLs with a total of five, whereas chromosomes 3, 6, and 10 each contained four QTLs. Chromosomes 2, 7, 8, and 9 each harbored three QTLs, while chromosome 4 had two, and chromosome 5 had only one QTL.

A total of 14 QTLs associated with GPC were detected in maize grain. Among these QTLs, six were identified at 30 DAP and located on chromosomes 3, 5, 6, 8, 9, and 10. Notably, qGPC30-6 exhibited the highest phenotypic variation, explaining 15.41% of variance, while the other five were small-effect QTLs. Notably, the favorable alleles of these five QTLs identified at 30 DAP originated from DH1M. Four QTLs were detected at 40 DAP, including one large-effect QTL (qGPC40-1) and three small-effect QTLs. Two QTLs were detected at 52 DAP, located on chromosomes 2 and 9, which explained 10.73% and 3.86% of the phenotypic variation respectively with all favorable alleles originating from DH1M. Additionally, two QTLs were detected at 61 DAP, located on chromosomes 3 and 8, with the highest PVE of 9.93% (Table 2 and Fig. 3). For GOC, twelve QTLs were identified, explaining between 2.91% and 12.23% of the total phenotypic variation. Only one QTL was associated with GOC at 30 DAP. However, six QTLs were detected at 40 DAP, with five carrying favorable allele originating from DH1M. At 40 DAP, GOC was controlled by two large-effect QTLs and four small-effect QTLs. At 52 DAP, three QTLs were identified for GOC, with qGOC52-6 explaining 14.62% of the total phenotypic variation. All favorable alleles of QTLs for GOC at 61 DAP were derived from DH1M. Regarding GSC, six QTLs were identified and explained 2.59%–19.39% of the total phenotypic variation. Two major QTLs, qGSC40-2 and qGSC52-4, explained 15.94% and 19.39% of the total phenotypic variation, respectively. Four QTLs carried the favorable allele originating from T877. Hardly any QTLs could be detected across different time points for GPC, GOC and GSC.

To further investigate the genetic basis of nutritional quality traits, we integrated the QTL mapping results with transcriptome and genome variation data to identify candidate genes associated with GPC, GOC, and GSC. A total of 4035 candidate genes were identified within the confidence interval of the QTLs (Fig. 4a, Table S1). Furthermore, multi-omics data were used to narrow down the list of candidate genes. In a previous study, we compared global gene expression of T877 and DH1M at 10, 20, 30, 40, and 50 DAP, detecting 1465 (DAP10) to 6980 (DAP40) DEGs [15,20]. Additionally, whole genome resequencing was conducted for T877 and DH1M, resulting in the annotation of 8598 nonsynonymous substitutions. By integrating QTL mapping, gene expression profiles, and comparative genomics data, we identified 157, 86, and 54 DEGs with nonsynonymous substitutions between parental lines associated with GPC, GOC, and GSC, respectively (Fig. 4a, Tables S2–S4). Subsequently, we performed gene function annotation, which resulted in the identification of 15 candidate genes potentially involved in regulating grain quality traits (Table 3).

GRMZM5G813651, the candidate gene for qGPC30-3 associated with GPC at 30 DAP (Table 3), encodes a NAC transcription factor. The expression level showed a significant difference between T877 and DH1M at 50 DAP, and one nonsynonymous SNP was observed in the third exon. The candidate genes for qGPC30-5 and qGPC30-6, GRMZM2G178753 and GRMZM2G474777, encode Leucine-rich repeat protein kinases. These genes showed significant differential expression between T877 and DH1M at 10 and 40 DAP, respectively (Fig. 4b and Table 3), with each containing a nonsynonymous SNP in exon 1. GRMZM5G814279, associated with qGPC30-9, encodes a MADS-box transcription factor. The expression level showed a significant difference between T877 and DH1M at 10 and 30 DAP, with a nonsynonymous SNP observed in the seventh exon. For qGPC30-10, GRMZM2G123633 encodes a cell wall invertase, showing significantly different between T877 and DH1M at 30 DAP. Additionally, BURP domain-containing protein GRMZM2G108422, and two transcription factors GRMZM2G073427 and GRMZM2G051256 were associated with GPC at 40 DAP. GRMZM2G169820, the candidate gene for qGOC40-1 associated with GOC at 40 DAP (Fig. 4b and Table 3), encodes an auxin response factor, and the expression level showed a significant difference between T877 and DH1M at 40 and 50 DAP. One nonsynonymous SNP was observed in the 13th exon of GRMZM2G169820. GRMZM2G173759 and GRMZM2G147966 were located in qGOC52-7 and encode Cellulose-synthase-like and Riboflavin synthase-like superfamily protein, respectively. GRMZM2G149994, the candidate gene for qGSC30-10 associated with GSC at 30 DAP (Fig. 4b and Table 3), is involved in the cell division cycle, and the expression level showed a significant difference between T877 and DH1M at 30 and 40 DAP. One nonsynonymous SNP was observed in the eighth exon of GRMZM2G149994. GRMZM2G162690, GRMZM2G060257 and GRMZM2G040743 were located in qGSC40-1 and encode trehalase, phloem protein 2-A13 and Calcium-dependent protein kinase family protein, respectively.