Genome-wide association study to identification of candidate genes influencing key seed traits in Perilla frutescens Authors: Me-Sun Kim, Gi-An Lee, Jung Yoon Yi, Yu-Mi Choi, Myoung-Jae Shin, Jae-Eun Lee, Sookyeong Lee, Kebede Taye Desta, Eun-Gyeong Kim, Gyeong Ju Gim, Eunae Yoo Published: 2025 ABSTRACT AbstractBackgroundPerilla frutescens is a widely cultivated annual crop in East Asia, valued for its oil-rich seeds and edible leaves. The seeds are abundant in functional compounds such as unsaturated fatty acids, polyphenols, and flavonoids, and oil-extraction by-products are used as protein-rich animal feed and organic fertilizer. Despite its economic and nutritional significance, the genetic basis underlying major seed traits in Perilla remains poorly understood.ResultsTo investigate the genetic architecture of key seed traits, a genome-wide association study (GWAS) was conducted using 61,539 high-quality SNPs derived from genotyping-by-sequencing (GBS) of 237 Perilla germplasms. Nine phenotypes were evaluated, including eight continuous traits (e.g., capsule length, seed height, thousand-grain weight) and one categorical trait (seed coat color). GWAS identified six significant SNPs distributed across five chromosomes, with two lead SNPs-one each on chromosomes 16 and 10-associated with thousand-grain weight and seed coat color, respectively. Haplotype analysis confirmed that these SNPs caused phenotypic differences among genotype groups. Importantly, two SNPs were located in the exons of candidate genes C2S51_032383 and C2S51_020378, both resulting in non-synonymous amino acid substitutions. Protein-protein interaction network analysis revealed associations with amino acid ligase family and cell wall organization pathways. Phylogenetic analysis further demonstrated that both genes are conserved among dicot species such as Arabidopsis thaliana and Glycine max.ConclusionsOur findings provide novel insights into the genetic determinants of seed-related traits in Perilla frutescens, especially seed weight and pigmentation. The trait-associated SNPs and candidate genes identified in this study offer promising targets for marker-assisted selection and molecular breeding aimed at improving seed quality and yield in Perilla.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12870-025-07714-0.Keywords: Perilla frutescens, Germplasm, Genome-wide association study, Thousand-grain weight, Seed coat color Background Perilla (Perilla frutescens) is an annual plant that belongs to the Lamiaceae family and is widely cultivated in Asian regions such as South Korea, China, and India [1, 2]. Perilla, one of the major oilseed crops, is used not only as food but also in industrial products such as paint and ink. The by-products generated after perilla seed oil extraction are utilized as animal feed or organic fertilizer [3]. This crop is attracting increasing attention for its health-promoting properties, including antioxidant activity, anticancer effects, and dementia prevention [4–6]. Perilla seeds contain approximately 40 to 50% fat, of which more than 90% consists of unsaturated fatty acids, including oleic acid (18:1), linoleic acid (18:2), and α-linolenic acid (18:3). These essential ω−3 fatty acids are known to be highly effective in lowering blood cholesterol levels and preventing various physiological disorders. In particular, linolenic acid, which accounts for more than 60% of perilla fatty acids, has been widely studied for its potential in preventing heart disease, alleviating allergies, and reducing cancer risk [7, 8]. Seed traits, including size, shape, seed coat thickness, and dormancy type, not only exhibit relatively stable heritability but are also closely associated with seed productivity, soil persistence, and germination success [9–13]. Seeds play a vital role in storing energy and nutrients to support early plant growth and development, while also enhancing offspring fitness, thereby facilitating population survival and establishment under adverse environmental conditions [14, 15]. In agriculture, seed traits serve as important indicators of genetic, physiological, and ecological factors that influence yield, quality, and market value. Breeding programs often focus on seed shape and size, with particular emphasis on size-related traits [16–19]. A comprehensive understanding of the genetic basis of seed morphology is essential for developing new cultivars with desirable agronomic characteristics. Recently, both linkage mapping and genome-wide approaches have been employed to dissect the complex genetic mechanisms underlying seed traits. Several genes and transcription factors involved in the morphological and anatomical regulation of seed traits have been identified in both monocotyledonous and dicotyledonous species [20–25]. According to Huang et al. [26], a high density haplotype-based genome-wide association study (GWAS) on 950 global rice varieties identified 32 novel loci associated with grain traits and flowering time, introducing a de novo assembly framework to uncover casual polymorphisms, while Priyanatha et al. [27] demonstrated the utility of GWAS in soybean by identifying five QTL regions and candidate genes linked to seed yield and quality traits in a diverse panel of 200 genotypes, highlighting the effectiveness of genomic tools in crop improvement. Despite increasing awareness of the nutritional and pharmacological benefits of Perilla frutescens, genomic research on this crop remains limited compared to other oilseed species. Although several studies have investigated fatty acid-related genes in perilla, little is known about the genetic regulation of its seed traits. In this study, we evaluated key seed-related traits using perilla germplasm collected from both domestic and international sources, and identified candidate genes associated with seed trait variation through GWAS. Materials and methods Plant materials and cultivation In this study, germplasms of perilla (n = 237) including 180 from Korea, 18 from China, 8 from Japan were obtained from the Gene bank of the National Agrobiodiversity Center, Rural Development Administration (RDA, Jeonju, Korea) (Supplementary Table 1), and were used as experimental materials (Supplementary Table 1). To evaluate phenotypic characteristics of 237 perilla germplasms, each plants were cultivated in an experimental field located at the center (latitude/longitude: 35o49 × 38.37´´N/127o09 × 07.78´´E) from the end of June to the end of November in 2023. At the end of June, seven seedlings of each germplasm were sown at a spacing of 60 cm in a row, and the rows were 30 cm apart. During the evaluation period from June to November 2023 (approximately six months), the monthly average temperature was recorded as 23.7 ℃, 27.1 ℃, 26.3 ℃, 22.3 ℃, 14.2 ℃, and 9.7 ℃, respectively, and the monthly average precipitation was recorded as 163 mm, 135 mm, 302 mm, 80 mm, 50.5 mm, and 67 mm. Fertilizers were sprayed in accordance with the manual standard cultivation method [28, 29]. Analysis of phenotypic characteristics Seed traits on 237 perilla germplasms were evaluated based on the manual on investigation of characteristics and management of genetic resources [30]. The flowering date and maturity period were observed, and the major seed traits were invested: the length of capsule (mm), width of capsule (mm), width of capsule entrance (mm), and ratio of capsule width. The length of capsule was measured middle part of apical inflorescence cluster based on main stem, and the width of capsule was measured the widest part of capsule. The width of capsule entrance was measured the most concave part of capsule entrance and the ratio of capsule width was measured the width of capsule entrance/the width of capsule. After harvest, the width (mm) and length (mm) of seeds were measured, and number of capsule per inflorescence cluster were invested. A thousand grain weight was measured with a regular commercial scale and the color of seed coat was measured by classifying it into seven types (Fig. 1). Schematic diagram of phenotypic investigation related ten key seed traits in perilla germplasms. A Investigation of capsule size by dividing into capsule length, capsule width, and capsule entrance width and calculation of ration of capsule width according to calculation formula; (B) Investigation of number of capsule per inflorescence cluster in main stem top apical part at maturity period; (C) Investigation of seed width and seed length; (D) Classification of seed coat color according to perilla seed coat color classification system; (E) Measurement of thousand grain weight Analysis of genotyping DNA extraction Genomic DNA was extracted from samples of young leaf tissue by using the NucleoSpin® Plant II kit (Macherey-Nagel™, Dueren, Germany) according to the manufacturer’s instructions. DNA quality and quantity were estimated by spectrophotometric measurements using a NanoDrop 3000 (Thermo Fisher Scientific, Waltham, MA, USA), with the A260/A280 ratio used as an indicator of protein contamination. The extracted DNA was diluted to concentration of 20 ng/ul and stored at −20 ℃ until further use. Genotyping and SNP calling Genotyping-by-Sequencing (GBS) was performed following a Elshire’ method [31] with slight modifications. The DNA libraries were generated by ApeKI restriction enzyme (New England Biolabs, Ipswich, MA, USA) and sequenced using an Illumina HiSeq X Ten platform system (Illumina, San Diego, CA, USA). After sequencing samples, shorted reads were trimmed using Trimmomatic [32] and aligned to the Perilla frutescens reference genome (version 1.0; https://ftp.ncbi.nlm.nih.gov/genomes/all/GCA/019/511/825/GCA_019511825.2_ICMM_Pfru_2.0/) [33] using BWA-MEN [34]. SAMtools [35] was used to generate BAM format and re-order reads. SNP calling was performed using HaplotypeCaller and CombineGVCFs [36]. Raw SNPs were filtered using Vcf tools [37] with following parameters: sequencing depth > 3, a minor allele frequency (MAF) > 0.01, and missing data < 30%. Genome-wide association study GWAS was conducted using the Genomic Association and Prediction Integrated (GAPIT) tool in R [38] utilizing the General Linear Models (GLM) [39] on following key seed traits: length of capsule, width of capsule, width of capsule entrance, ratio of capsule width, number of capsule per inflorescence cluster, width of seed, length of seed, thousand grain weight, color of seed coat. To reduce the risk of false positives arising from population structure and kinship, additional analyses were carried out using four alternative models: Mixed Linear Model (MLM), Multiple Loci Mixed Model (MLMM), Fixed and Random Model Circulating Probability Unification (FarmCPU), and Bayesian-information and Linkage-disequilibrium Iteratively Nested Keyway (BLINK) [40–42]. Cross-validation was performed by comparing SNP associations across the five models, and only associations consistently detected in multiple models were considered robust. Genome-wide Manhattan plots and multitrack quantile-quantile (Q-Q) plots were visualized with the qqman package of R [43]. Significant SNPs were defined based on a false discovery rate (FDR) threshold of < 0.05, and results were compared with Bonferroni-adjusted thresholds for stringent validation [44]. Identification of candidate genes Linkage disequilibrium (LD) calculations were performed using TASSEL software (version 5.0) [45], considering only SNPs with MAF > 0.01 to avoid bias effects on LD. Haplotype blocks were constructed based on genotypes on LD statistics r2 values within 100 kb gnomic regions surrounding significant SNPs. Haplotype block visualization was performed using SRplot, a free online platform for data visualization and graph creation (http://www.bioinformatics.com.cn/srplot) [46]. Statistical analysis was performed using the ggpubr package in R to compare perilla germplasm groups with different haplotypes. Candidate genes in the same position as significant SNPs were annotated based on the reference genome of Perilla frutescens (GCA_019511825.2_ICMM_Pfru_2.0). Their molecular functions and amino acid compositions were analyzed using the NCBI database (http://www.ncbi.nlm.nih.gov) and the ProtParam tool (http://www.expasy.org), and the predicted tertiary protein structures were generated using SWISS-MODEL (http://www.swissmodel.expasy.org/interactive). In addition, a protein-protein interaction network was constructed using the STRING database (https://cn.string-db.org) to visualize the functional relationships between candidate gene proteins and interacting proteins. Phylogenetic analysis Protein sequence of the candidate gene, as annotated in Perilla frutescens reference genome (GCA_019511825.2_ICMM_Pfru_2.0), was blasted against ten protein databases each of monocots and dicotyledons (https://www.ncbi.nlm.nih.gov/BLAST.cgi/) using BLASTp. The top homologous sequence with the highest similarity scores was retrieved and subjected to multiple sequence alignment. A phylogenetic tree was subsequently constructed based on these aligned sequences using the neighbor-joining (NJ) method implemented in MEGA-X [47], with 1,000 bootstrap replicates to assess branch support. Expression analysis of soybean orthologs Orthologs of P. frutescens candidate genes (C2S51_032383 and C2S51_020379) were identified as Glyma.05G214300 and Glyma.15G223800 using BLASTP against the soybean reference genome (Glycine max Wm82.a4.v1). Tissue-specific expression profiles of these orthologs were obtained from the Phytozome database (https://phytozome-next.jgi.doe.gov/). Normalized expression values (FPKM) across six tissues (leaf, stem, shoot, root, flower, and seed) were extracted and visualized using R software. Statistical analysis A one-way analysis of variance (ANOVA) was performed to analyze the phenotypic data. Frequency distributions, box plots, and correlation between key seed traits were analyzed and visualized using the ggplot2 package and the SNP density plots were visualized using the CMplot package in RStudio (version 4.4.2) [48]. Results Phenotypic evaluation among perilla germplasm To investigate the variation of key seed traits in 237 perilla germplasms that grow normally in accordance with domestic environments, the nine phenotypic traits was investigated during approximately six months in 2023. These traits could be divided into three groups; capsule-related traits (capsule length, capsule width, capsule entrance width, ratio of capsule width), seed-related traits (seed length, seed width, color of seed coat, thousand grain weight), and inflorescence cluster-related traits (number of capsule per inflorescence cluster). The phenotypic evaluation revealed a normal distribution in genetic parameters for eight key seed traits except color of seed coat (Supplementary Fig. 1). The standard deviation value was highest for number of capsule per inflorescence cluster (10.583) and lowest for ratio of capsule width (0.082) among 8 continuous key seed traits except level-specific trait, color of seed coat. Among the evaluated 237 perilla germplasms, the highest values for capsule-related and seed-related traits were observed in a landrace of unknown origin (IT214486) and the highest number of capsule per inflorescence cluster was observed in a Chinese landrace (ITK248224). For the color of seed coat, level-2 showed the highest frequency, comprising 107 germplasms, whereas level-7 included only one perilla germplasm (IT261878, Korea origin breeding line) (Supplementary Fig. 2 and Supplementary Table 2). Several significant correlations were observed among nine capsule-related and seed-related traits analyzed in perilla germplasms (Fig. 2 and Supplementary Fig. 3). A strong positive correlation was detected between capsule length (CL) and capsule width (CW) (r = 0.730, ***), as well as between capsule width (CW) and capsule entrance width (CEW) (r = 0.748, ***), suggesting that larger capsules tend to have proportionally wider entrances. These associations may reflect shared developmental mechanisms governing overall capsule size. In seed-related traits, seed width (SW), seed length (SL), and thousand grain weight (TGW) showed highly significant and positive correlations, with SW–SL exhibiting an almost perfect correlation (r = 0.993, ***). Both SW-TGW (r = 0.695, ***) and SL-TGW (r = 0.686, ***) correlations imply that seed dimensions are major contributors to seed weight. These results indicate that selection for larger seed size may simultaneously improve seed mass. The number of capsules per inflorescence cluster (NCF) displayed overall negative correlations with most other traits, especially CEW (r = −0.298, ***) and SW (r = −0.213, ***), suggesting a possible trade-off between reproductive output and individual seed development. This trend supports insights that increased capsule production may occur at the expense of seed size or mass. Lastly, color of seed coat (CSC) was moderately but significantly correlated with seed width (r = 0.376, ***), seed length (r = 0.357, ***), and thousand grain weight (r = 0.388, ***). These results indicate that brighter seed coats may be associated with larger and heavier seeds, possibly due to pleiotropic genetic control or linkage among seed pigmentation and development genes. Distribution and correlation analysis of estimated phenotypic values for perilla seed traits. Asterisks indicate significant Pearson’s correlations, n = 237 perilla germplasms. CL, Capsule length; CW, Capsule width; CEW, Capsule entrance width; RCW, Ration of capsule width; NCF, Number of capsule per inflorescence cluster; SW, Seed width; SL, Seed length; TGW, Thousand-grain weight; CSC, Color of seed coat Genotyping and SNP calling A total of 723,274 SNPs were called after trimming and mapping against the Perilla frutescence reference genome (version 1.0; GCA_019511825.2_ICMM_Pfru_2.0/). A final number of 61,539 SNPs distributed across all twenty chromosomes of the perilla genome were identified after filtration (Fig. 3). The chromosomal distribution of SNPs ranged from 4794 SNPs on Chr.5 to 1505 SNPs on Chr.20 (Supplementary Table 3). Additionally, the SNP frequency was highest on Chr.5 with 3.83 SNPs per 50 kb. PCA based on the genotype data of 237 perilla germplasms revealed that the first two principal components (PC1 and PC2) explained 36.1% and 4.7% of total genetic variation, respectively. This indicates that the substantial portion of total genomic variance was captured by the top two components. Each germplasms was partially clustered under PC1 and PC2, and some individuals showed a dispersed distribution pattern. In the kinship matrix, strong genetic similarity (indicated in red) was observed among certain groups of accessions, suggesting close genetic relationship (Supplementary Fig. 4). SNP density plot on 20 perilla chromosome showing number of SNPs within 1 Mb window size GWAS analysis of key seed traits To identify genomic loci associated with the key seed traits, GWAS was conducted using 61,539 SNPs across all nine phenotypic traits. The analysis was performed using GLM model in GAPIT package, generating both Q-Q plots and Manhattan plots (Fig. 4). The Q-Q plots exhibited a clear upward deviation in tail region from expected distribution, indicating that both false positives and false negatives were effectively controlled in model. A set of significant SNPs that were associated with key seed traits was identified according to the set threshold of Manhattan plots. Based on the Manhattan plots for nine key seed traits, significant SNPs were identified for six traits (capsule width, number of capsule per inflorescence cluster, seed width, seed length, thousand grain weight, and color of seed coat) according to threshold criteria. The highest number of significant SNPs (a total of 15 SNPs) was detected for seed length, while the highest peak with the strongest statistical significance (p-value of 2.82E-16) was observed for color of seed coat and ranged from 8.6 to 15.5 corresponding –log10(p-value). In the Manhattan plot of thousand grain weight, significant SNPs were distributed across eight chromosomes. Notably, the GWAS results for seed width and seed length exhibited similar patterns of SNP detection, suggesting a potential genetic correlation between these traits. Manhattan plots and Q-Q plots showing significantly associated SNPs detected in combined environment GWAS analysis for nine key perilla seed related traits. A Capsule length (CL); B Capsule width (CW); C Capsule entrance width (CEW); D Ration of capsule width (RCW); E Number of capsule per inflorescence cluster (NCF); F Seed width (SW); G Seed length (SL); H Thousand-grain weight (TGW); I Color of seed coat (CSC) Identification of candidate genes affecting key seed traits in perilla germplasms Haplotype blocks were examined within ± 100 kb region surrounding SNPs showing strong trait associations. Among the six traits with highly significance in the Manhattan plots, four (thousand grain weight, seed width, seed length, and color of seed coat) exhibited clear LD block formation encompassing the lead SNPs. For thousand grain weight, relevant loci were located on chromosomes 9 (Supplementary Fig. 5), 16 (Fig. 5), and 19 (Supplementary Fig. 6). Additionally, regions associated with seed size and seed coat color were located to chromosomes 14 (Supplementary Fig. 7 and Supplementary Fig. 8) and 10 (Fig. 6), respectively. For thousand grain weight, six SNPs within a high LD block (r2 = 0.938) on chromosome formed two distinct haplotype groups. Comparative analysis of TGW across these haplotypes revealed statistically significant differences among the groups. On chromosome 14, LD blocks harboring seed size-associated lead SNPs (S14-55869358 and S14-15396999) revealed the presence of two and three haplotype groups, respectively, which were also significantly associated with seed size variation. Regarding color of seed coat, an LD block on chromosome 10 (r2 = 0.942) contained four SNPs forming five haplotype groups. Evaluation of seed coat color distribution and germplasm frequency across these haplotypes showed that Hap2 included the largest number of germplasms, while Hap3 predominantly contained perilla germplasms with light-colored seed coats (Level9 and Level10) (Fig. 6). Furthermore, annotation of the lead SNPs identified through GWAS and LD analysis using perilla genome database revealed that those associated with thousand grain weight and seed coat color were located within coding regions of annotated genes. Based on this, structural and functional analyses were conducted on two lead SNPs to further explore their potential biological relevance. Identification of significantly SNP (S16_19291606) and haplotype related to thousand-grain weight. A Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 16; B Details and frequency of two haplotype among 237 perilla germplasms; C Comparison of thousand-grain weight(g) between different haplotypes using Tukey’s test Identification of significantly SNP (S10_31019666) and haplotype related to color of seed coat. A Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 10; B Details and frequency of five haplotype among 237 perilla germplasms; C Comparison of seed coat color distribution according to different haplotypes; D, Color of seed coat phenotypes of perilla seeds according to different haplotypes. Hap1-Level3, IT322029; Hap2-Level1, IT208792; Hap2-Level5, IT216931; Hap3-Level1, IT183207; Hap3-Level9, IT208847 Identification of structural and functional modification To identify potential candidate genes associated with key seed traits, a genome-wide search was conducted using the perilla genome database, focusing on lead SNPs that showed strong associations with traits. Previously, GWAS using the GLM model identified two significant SNPs, S16_19291606, associated with thousand-grain weight, and S10_31019666, associated with color of seed coat. To verify robustness of these associations, additional cross-validation analyses were performed using four alternative GWAS models (MLM, MLMM, FarmCPU, and BLINK) (Supplementary Fig. 9 and supplementary Fig. 10). Both loci were consistently detected as the strongest association peaks across all five models, indicating that identified signals were not artifacts of population structure or model bias. To further investigate the genetic basis of these associations, all annotated genes were extracted within ± 100 kb of lead SNPs. For the locus associated with color for seed coat, a total of 12 candidate genes, including C2S51_020378, were identified. For the locus associated with thousand-grain weight, one candidate gene, C2S51_032383, was found to be located within the interval. A detailed list of genes and their annotations are provided in Supplementary Table 6. Among these, two genes, C2S51_032383 and C2S51_020378 were identified with lead SNPs located within exon regions (Fig. 7). C2S51_032383 which consists of three exons encodes a protein composed of 684 amino acids. The lead SNP (S16_19291610) was located in third exon at nucleotide position 2,256. A point mutation at this position resulted in an amino acid substitution from glycine to cysteine at amino acid position 532, affecting the structural conformation of protein (Fig. 7B-a). C2S51_020378 comprises three exons and encodes a protein of 549 amino acids. The lead SNP, S10_31019666 (g.141G > A) was located in first exon and caused a non-synonymous substitution of valine to methionine at amino acid position 47, indicating a nucleotide substitution within coding sequence (Fig. 7B-b). To further investigate biological significance of these SNP-induced structural changes, protein-protein interaction networks involving C2S51_032383 and C2S51_020378 were analyzed, which allowed for the exploration on their functional relationship with other interacting proteins. Localization of SNP and prediction model for protein structure changes according to SNP mutation. A Structure of two candidate genes (C2S51_032383 and C2S51_020378) and corresponding SNPs related to thousand-grain weight and seed coat color, respectively; B-a, Comparison of wild type C2S51_032383 protein structure (right) with its mutant forms (left) having mutation from Glycine to Cysteine at position 532; B-b, Comparison of wild type C2S51_020378 protein structure (right) with its mutant forms (left) having mutation from Valine to Methionine at position 47 Functional characterization of candidate genes through protein interaction and phylogenetic tree analysis To further investigate biological roles of two candidate genes identified through GWAS (C2S51_032383 and C2S51_020378), functional annotations were conducted using protein-protein interaction network and phylogenetic analysis based on protein sequence homology. In protein interaction analysis, protein sequences of both candidate genes were queried against STRING database to predict functional interaction networks (Fig. 8). C2S51_032383 was found to interact with proteins involved in three major functional categories: amino acid ligase family, cellular water homeostasis, and glucanotransferase activity. Amino acid ligases are ATP-dependent enzymes essential for peptide bond formation and storage protein biosynthesis during seed maturation. Their interaction with C2S51_032383 suggests a role in promoting nutrient accumulation and protein deposition, potentially enhancing seed weight. The association with aquaporin-like proteins and other water homeostasis regulators implies a possible involvement in maintaining intracellular hydration, which may affect seed filling and density. Additionally, glucanotransferase interactions support a potential role in modifying cell wall polysaccharides, which are critical for seed structure and development. C2S51_020378 showed interactions with methylesterase family proteins, extensin-related proteins, and structural components of the cell wall. Methylesterases are known to modulate pectin properties, influencing cell wall rigidity and permeability, while extensins (hydroxyproline-rich glycoproteins) are involved in reinforcing the cell wall during development and under stress conditions. These associations imply a functional role for C2S51_020378 in shaping seed coat characteristics through regulation of wall structure, pigment deposition, and response to environmental stimuli. To explore evolutionary conservation, phylogenetic analyses were conducted using BLAST-based alignments of both candidate genes against orthologs from twelve representative plant species, including five monocots, five dicots, and two Lamiaceae family members (Fig. 9). For C2S51_032383, a gene associated with thousand grain weight, P. frutescens formed a close clade with Solanum lycopersicum with high bootstrap support (100), indicating strong conservation. This clade was further linked to Scutellaria baicalensis, a fellow Lamiaceae species, suggesting potential functional similarity within the family. More distantly related were monocot species such as Oryza sativa and Triticum aestivum, which showed lower sequence similarity. These results suggest that the gene’s role in seed biomass accumulation may be conserved primarily among dicot species, particularly within Solanaceae and Lamiaceae. In contrast, C2S51_020378, which was associated with seed coat color, clustered most closely with Solanum tuberosum (bootstrap 69) and shared a broader clade with Glycine max, Arabidopsis thaliana, and Brassica napus. Notably, Scutellaria baicalensis appeared closer to S. lycopersicum than to Perilla, and Coriandrum sativum was placed on a distinct branch, suggesting variable divergence even within the same botanical family. The separation between dicots and monocots was clearly delineated, with monocotyledonous species forming a well-supported cluster apart from the Perilla candidate genes. Overall, both candidate genes demonstrated higher sequence similarity to other dicot proteins than to monocots, supporting their functional conservation across dicot lineages. However, the unexpected distances observed within Lamiaceae species, particularly for C2S51_020378, highlight possible gene-specific evolutionary divergence. These patterns may reflect differences in selective pressures acting on genes related to seed size and seed coat traits across species. Protein-protein interaction network based on STRING database of two perilla genes, C2S51_032383 (A) and C2S51_020378 (B). Line color indicates the type of interactions. Black, gene of co-expression; Green, gene of neighborhood; Blue, gene of co-occurrence; Pink, experimentally determined gene. Line thickness indicates the strength of data support Phylogenetic tree analysis of two Perilla gene, C2S51_032383 (A) and C2S51_020378 (B) based on protein sequences across 12 plant species. Phylogenetic trees were constructed using amino acid sequences of candidate protein and its homologs identified via BLASTP against five monocotyledonous species (Oryza sativa, Zea mays, Sorghum bicolor, Hordeum vulgare, Triticum aestivum), five dicotyledonous species (Arabidopsis thaliana, Glycine max, Solanum tuberosum, Solanum lycopersicum, Brassica rapa), and two members of the Lamiaceae family (Coriandrum sativum, Scutellaria baicalensis). Multiple sequence alignment was performed using ClustalW, and tree construction was carried out in MEGA-X using the Neighbor-Joining method with 1,000 bootstrap replicates Discussion In this study, we conducted a phenotypic analysis of major seed traits and GWAS based on GBS in 237 perilla germplasms adapted to external environments. Our aim was to identify key seed trait-associated SNPs and candidate genes. Most continuous traits showed normal distribution, and the highest phenotypic variation was observed in the number of capsule per inflorescence cluster. This indicates a high degree of genetic diversity within perilla germplasm, which may be valuable for breeding purposes. This result aligns with previous reports on perilla genetic resources [49, 50], where substantial genetic variation in seed-related traits, particularly thousand grain weight and seed size, was also observed. Park et al. [51] analyzed genetic variation of seed characteristics and fatty acid composition using simple sequence repeats (SSR) markers for 280 traditional perilla germplasms in Korea. The study confirmed significant variation in seed size, hardness, and major fatty acid content, suggesting that it is an important genetic germplasm for the development of perilla varieties. A total of 61,539 high-quality SNPs were identified, and population structure analysis using PCA and kinship matrix revealed clusters with high genetic similarity. These findings suggest the presence of distinct genetic backgrounds among landraces, breeding lines, and wild types, consistent with prior studies on genetic structure of perilla [52]. The GWAS analysis identified six significant SNPs associated with key seed traits, and similar SNP distribution patterns were observed for seed length and width, supporting presence of genetic correlation between these traits. These results are consistent with GWAS results observed in other species such as Arabidopsis [53], soybean [54, 55], mungbean [56], etc. [57, 58], where seed size-related traits were shown to be governed by shared regulatory networks. Notably, significant SNPs associated with thousand grain weight were found on chromosomes 9, 16, and 19, with major SNPs located within functionally relevant genes inside LD blocks. For instance, C2S51_032383 interacts with amino acid ligases involved in seed storage protein biosynthesis and aquaporins related to intracellular water homeostasis. These interactions resemble those reported in rice during grain filling stage, where aquaporins and sucrose transporters play crucial roles [59, 60]. The linkage between thousand grain weight and genes associated with storage protein biosynthesis suggests that such genes could serve as effective markers for selection in breeding programs aimed at improving seed weight [61, 62]. Notably, C2S51_032383 was the only candidate gene identified within the ± 100 kb region surrounding significant lead SNP associated with thousand grain weight. Expression profiling based on public RNA-seq data (Glycine max Wm82.a4.v1, Phytozome) further supports its functional role. Orthologous genes in soybean (Glyma.05G214300 and Glyma.15G223800) exhibited detectable expression across multiple tissues, including leaf, stem, shoot, root, flower, and seed (Supplementary Fig. 11). Importantly, in seed developmental stages (stages 1–9), expression of Glyma.05G214300 was maintained at moderate levels during the early stages (stages 1–3), when seed filling and storage compound accumulation are initiated. This temporal expression pattern suggests that C2S51_032383 may contribute to regulation of early seed development and biomass deposition, consistent with its predicted role in amino acid ligase activity and interaction with aquaporin-mediated water homeostasis. Such evidence from soybean, a closely related oilseed crop, provides cross-species validation and highlights conserved functional significance of this gene in determining seed weight. These cross-species parallels reinforce that the genetic architecture underlying seed weight and pigmentation traits is at least partly conserved among oilseed crops. For instance, genetic findings on oil production and quality in sesame have identified multiple loci for oil content, fatty acid composition, and yield, showing that seed weight and oil-related genes are concentrated in regions affecting lipid metabolism [63]. Similarly, gene expression profiles that shape high and low oil content sesames showed that lipid biosynthesis genes such as DGAT, PDAT, and GPAT are differentially expressed in seed development stages between high- and low‐oil varieties, especially during mid to late seed filling [64]. These studies underscore that genes involved in storage protein biosynthesis and water homeostasis, similar to our candidate C2S51_032383, are part of conserved pathways regulating seed weight and composition in oilseed crops. The seed coat color trait showed the highest frequency at level-2 pigmentation, and germplasms with bright seed coat colors (levels 9 and 10) were mainly found in a specific haplotype group (Hap3). C2S51_020378 interacts with pectin methylesterase and extensin proteins related to cell wall architecture. This functional role is similar to genes identified in Arabidopsis [65–67] and pumpkin accession [68, 69], where cell wall re-modeling proteins influence seed coat coloration. In Brassica napus, a rare dominant allele DYSOC1 was found to determine seed coat color and improve seed oil content, highlighting the role of specific genes in seed coat pigmentation and quality [70]. Moreover, studies in sesame have identified genes associated with seed coat color, such as WRKY transcription factor 23 and salicylic acid-binding protein 2, demonstrating the conserved nature of seed coat color regulation across species [71]. The LD block on chromosome 10 containing this gene included four significant SNPs and five distinct haplotype groups with high polymorphism, indicating a complex genetic regulation of this trait. The phylogenetic analysis of two candidate genes, C2S51_032383 and C2S51_020378, was expanded to include not only monocot and dicot model plants but also two additional members of the Lamiaceae family (Coriandrum sativum and Scutellaria baicalensis). This inclusion was motivated by taxonomic proximity of these species to Perilla frutescens, allowing for a more nuanced understanding of lineage-specific gene conservation. The phylogenetic trees demonstrated that P. frutescens grouped closely with both C. sativum and S. baicalensis, indicating a strong evolutionary relationship within the Lamiaceae clade. This observation aligns with hypothesis that certain regulatory mechanisms, especially those involved in seed development, pigmentation, and structural integrity, may be conserved within this plant family. Specifically, C2S51_032383 showed strong sequence homology with genes encoding members of the amino acid ligase family. These ligases are known to catalyze peptide bond formation in an ATP-dependent manner and are essential for the biosynthesis of seed storage proteins during maturation. Previous studies in Setaria italica demonstrated that overexpression of an E3 ubiquitin ligase (SGD1) significantly increased grain yield and seed weight, underscoring the functional role of ligase family proteins in regulating biomass accumulation [72]. Similarly, in Arabidopsis, the DA1 and DA2 genes, both coding for ubiquitin ligases, are directly involved in controlling seed and organ size [73]. Furthermore, its interaction with aquaporin-related proteins is particularly noteworthy. Aquaporins are membrane proteins that facilitate water transport and are highly expressed during seed development. Their role in maintaining cellular hydration and supporting seed filling has been well-documented in crops such as maize and soybean [74, 75], suggesting that C2S51_032383 may also contribute to thousand grain weight via coordinated control of water homeostasis and protein accumulation. In contrast, C2S51_020378 was phylogenetically related to genes from methylesterase and extensin protein families, both of which are involved in modifying the structural integrity of seed coat cell wall. Methylesterase enzymes (PMEs) catalyze the removal of methyl groups from pectin, altering the rigidity, porosity, and adhesion of cell wall matrix. Studies in Arabidopsis have shown that mutants in PME-related genes such as PMEI6 and PMEI14 exhibit altered mucilage secretion and seed coat architecture, indicating their key role in modulating cell wall structure during seed development [76, 77]. While these effects are not direct indicators of pigment biosynthesis, they are closely linked to the physical environment in which pigments are deposited and stabilized. Likewise, extensin proteins are hydroxyproline-rich glycoproteins that reinforce the plant cell wall under both developmental and stress-induced conditions. In soybean, extensins are localized to specific layers of the seed coat, influencing its structural robustness and potentially affecting traits like pigment retention and permeability [78]. Together, these findings indicate that C2S51_032383 and C2S51_020378 are not only phylogenetically conserved across diverse angiosperms but may also perform functionally analogous roles in seed trait regulation. Furthermore, a negative correlation was observed between the number of capsules per inflorescence cluster and most other traits. This suggests a reproductive trade-off, in which increased floral organ development may lead to reduced seed size or weight due to resource allocation limitations. Similar reproductive trade-offs have been reported in soybean and other crops [79, 80], and our findings support that such mechanisms may also operate in perilla. Overall, this study identified trait-associated SNPs and candidate genes related to major seed characteristics in Perilla frutescens germplasms and explored their potential functional interactions. These findings not only contribute to a better understanding of the genetic architecture underlying seed morphology and productivity traits but also underscore the considerable genetic diversity preserved in traditional landraces and wild relatives. This diversity represents a crucial reservoir of alleles that can be harnessed for future breeding programs, particularly under the challenges of climate variability and evolving agricultural demands. By integrating these genomic insights into marker-assisted selection (MAS) and molecular breeding strategies, this research provides a robust foundation for accelerating trait improvement and developing high-yielding, stress-resistant perilla cultivars tailored to specific agronomic goals. The results further highlight the strategic importance of conserving and utilizing perilla genetic resources to ensure food and nutritional security, particularly in regions where perilla plays a culturally and economically significant role. The integration of GWAS, haplotype analysis, and protein-protein interaction network analysis provides functional support for the genetic associations identified in this study, setting it apart from previous work on perilla germplasms. To further elucidate the molecular mechanisms underlying seed trait variation, future research should incorporate RNA-seq-based transcriptome profiling across key developmental stages, especially focusing on differentially expressed genes between contrasting haplotypes. Additionally, CRISPR/Cas-mediated gene editing can serve as a powerful tool for functional validation of candidate genes, enabling the establishment of causal relationships between genetic variation and phenotype. Beyond validation, metabolomic and epigenomic analyses may offer complementary insights into regulatory layers affecting seed development and oil biosynthesis. Such integrative multi-omics approaches will be instrumental in accelerating marker-assisted selection and genomic selection strategies for perilla breeding programs aimed at yield improvement, quality enhancement, and stress resistance. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary figure 1. Distribution of nine key seed related traits among 237 Perilla frutescens germplasms. A, Capsule length (CL); B, Capsule width (CW); C, Capsule entrance width (CEW); D, Ratio of capsule width (RCW); E, Number of capsule per inflorescence cluster (NCF); F, Seed width (SW); G, Seed length (SL); H, Thousand-grain weight (TGW); I, Color of seed coat (CSC). Supplementary figure 2. Box plots showing the phenotypic values of nine key seed traits. The black horizontal lines in middle of the boxes are the median values. A, Capsule length (CL); B, Capsule width (CW); C, Capsule entrance width (CEW); D, Ratio of capsule width (RCW); E, Number of capsule per inflorescence cluster (NCF); F, Seed width (SW); G, Seed length (SL); H, Thousand-grain weight (TGW). Supplementary figure 3. Principal component analysis (PCA) and biplot vectors of nine key seed traits in the 237 perilla germplasms. CL, Capsule length; CW, Capsule width; CEW, Capsule entrance width; RCW, Ratio of capsule width; NCF, Number of capsule per inflorescence cluster; SW, Seed width; SL, Seed length; TGW, Thousand-grain weight; CSC, Color of seed coat. Supplementary figure 4. PCA and Kinship matrix of 237 Perilla frutescens germplasms. A, PCA eigenvalue plot showing the number of principal components versus the percentage of variance explained by each component; B, PCA of 237 perilla germplasms based on the first two principal components. Each data point represents a single perilla germplasms; C, Heatmap of the kinship matrix, which represents the degree of relatedness between each pair of perilla germplasms in dataset. Supplementary figure 5. Identification of significantly SNP (S9_58899221) and haplotype related to thousand-grain weight. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 9; B, Details and frequency of two haplotype among 237 perilla germplasms; C, Comparison of thousand-grain weight(g) between different haplotypes using Tukey’s test. Supplementary figure 6. Identification of significantly SNP (S19_14145084) and haplotype related to thousand-grain weight. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 19; B, Details and frequency of three haplotype among 237 perilla germplasms; C, Comparison of thousand-grain weight(g) between different haplotypes using Tukey’s test. Supplementary figure 7. Identification of significantly SNP (S14_55869358) and haplotype related to seed width and length. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 14; B, Details and frequency of two haplotype among 237 perilla germplasms; C, Comparison of seed width(mm) between different haplotypes using Tukey’s test. Supplementary figure 8. Identification of significantly SNP (S14_15396999) and haplotype related to seed width and length. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 14; B, Details and frequency of three haplotype among 237 perilla germplasms; C, Comparison of seed width(mm) between different haplotypes using Tukey’s test. Supplementary figure 9. Circular Manhattan plot and Q-Q plots using five GWAS models for cross-validation of thousand-grain weight in perilla germplasms. Manhattan plots showing significantly associated SNPs detected in combined environment GWAS analysis. Circular Manhattan plots sequentially represents BLINK, FarmCPU, MLMM, MLM, and GLM from the outer to inner circles. Supplementary figure 10. Circular Manhattan plot and Q-Q plots using five GWAS models for cross-validation of color of seed coat in perilla germplasms. Manhattan plots showing significantly associated SNPs detected in combined environment GWAS analysis. Circular Manhattan plots sequentially represents BLINK, FarmCPU, MLMM, MLM, and GLM from the outer to inner circles. Supplementary figure 11. Tissue-specific gene expression levels of soybean (Glycine max) orthologs of P. frutescens candidate genes. Expression levels of Glyma.05G214300 and Glyma.15G223800 (orthologs of C2S51_032383 and C2S51_020379) across leaf, stem, shoot, root, flower, and seed, based on publicly available RNA-seq datasets. Supplementary Material 5. Authors’ contributions M-S.K. conceived and coordinated the study, performed formal analysis and visualization, and wrote the original draft. G-A.L., J.Y.Y. and Y-M.C. provided resources and reviewed the manuscript. M-J.S., J-E.L., S.L., K.T.D., E-G.K. and G.J.G reviewed the manuscript. E.Y. acquired funding, managed the project, conceptualized the study, developed the methodology, conducted the investigation. Funding This study was carried out with the support of Research Program for Agriculture Science and Technology Development (Project No. PJ01676603 & PJ01619101), Rural Development Administration, Republic of Korea and supported by the 2025 RDA Fellowship Program of the National Institute of Agricultural Sciences, Rural Development Administration, Republic of Korea. Data availability The datasets created and/or analyzed during the current investigation are accessible from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate Not applicable Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary figure 1. Distribution of nine key seed related traits among 237 Perilla frutescens germplasms. A, Capsule length (CL); B, Capsule width (CW); C, Capsule entrance width (CEW); D, Ratio of capsule width (RCW); E, Number of capsule per inflorescence cluster (NCF); F, Seed width (SW); G, Seed length (SL); H, Thousand-grain weight (TGW); I, Color of seed coat (CSC). Supplementary figure 2. Box plots showing the phenotypic values of nine key seed traits. The black horizontal lines in middle of the boxes are the median values. A, Capsule length (CL); B, Capsule width (CW); C, Capsule entrance width (CEW); D, Ratio of capsule width (RCW); E, Number of capsule per inflorescence cluster (NCF); F, Seed width (SW); G, Seed length (SL); H, Thousand-grain weight (TGW). Supplementary figure 3. Principal component analysis (PCA) and biplot vectors of nine key seed traits in the 237 perilla germplasms. CL, Capsule length; CW, Capsule width; CEW, Capsule entrance width; RCW, Ratio of capsule width; NCF, Number of capsule per inflorescence cluster; SW, Seed width; SL, Seed length; TGW, Thousand-grain weight; CSC, Color of seed coat. Supplementary figure 4. PCA and Kinship matrix of 237 Perilla frutescens germplasms. A, PCA eigenvalue plot showing the number of principal components versus the percentage of variance explained by each component; B, PCA of 237 perilla germplasms based on the first two principal components. Each data point represents a single perilla germplasms; C, Heatmap of the kinship matrix, which represents the degree of relatedness between each pair of perilla germplasms in dataset. Supplementary figure 5. Identification of significantly SNP (S9_58899221) and haplotype related to thousand-grain weight. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 9; B, Details and frequency of two haplotype among 237 perilla germplasms; C, Comparison of thousand-grain weight(g) between different haplotypes using Tukey’s test. Supplementary figure 6. Identification of significantly SNP (S19_14145084) and haplotype related to thousand-grain weight. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 19; B, Details and frequency of three haplotype among 237 perilla germplasms; C, Comparison of thousand-grain weight(g) between different haplotypes using Tukey’s test. Supplementary figure 7. Identification of significantly SNP (S14_55869358) and haplotype related to seed width and length. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 14; B, Details and frequency of two haplotype among 237 perilla germplasms; C, Comparison of seed width(mm) between different haplotypes using Tukey’s test. Supplementary figure 8. Identification of significantly SNP (S14_15396999) and haplotype related to seed width and length. A, Chromosome-wise Manhattan plot of single-polymorphism-based association and linkage disequilibrium (LD) heatmap surrounding the peak on chromosome 14; B, Details and frequency of three haplotype among 237 perilla germplasms; C, Comparison of seed width(mm) between different haplotypes using Tukey’s test. Supplementary figure 9. Circular Manhattan plot and Q-Q plots using five GWAS models for cross-validation of thousand-grain weight in perilla germplasms. Manhattan plots showing significantly associated SNPs detected in combined environment GWAS analysis. Circular Manhattan plots sequentially represents BLINK, FarmCPU, MLMM, MLM, and GLM from the outer to inner circles. Supplementary figure 10. Circular Manhattan plot and Q-Q plots using five GWAS models for cross-validation of color of seed coat in perilla germplasms. Manhattan plots showing significantly associated SNPs detected in combined environment GWAS analysis. Circular Manhattan plots sequentially represents BLINK, FarmCPU, MLMM, MLM, and GLM from the outer to inner circles. Supplementary figure 11. Tissue-specific gene expression levels of soybean (Glycine max) orthologs of P. frutescens candidate genes. Expression levels of Glyma.05G214300 and Glyma.15G223800 (orthologs of C2S51_032383 and C2S51_020379) across leaf, stem, shoot, root, flower, and seed, based on publicly available RNA-seq datasets. Supplementary Material 5. Data Availability Statement The datasets created and/or analyzed during the current investigation are accessible from the corresponding author upon reasonable request.