Large-scale GWAS integrated with image-based phenotyping reveals loci and trait-associated markers for Perilla frutescens seed traits

Abstract

Background

Perilla (Perilla frutescens) is an important oilseed crop in East Asia with high nutritional and economic value. Seed size and seed coat color are key agronomic traits influencing yield, oil quality, and market preference. However, genetic studies in perilla remain limited by small population sizes and low-throughput phenotyping, restricting their application in breeding.

Results

A large-scale genome-wide association study (GWAS) was conducted by integrating high-throughput image-based phenotyping in a panel of 493 perilla accessions. Measurements of seed morphology, including area, perimeter, and length, and color traits (RGB components) were obtained. Genotyping-by-sequencing generated high-quality single-nucleotide polymorphism (SNP) datasets, and GWAS was performed using four statistical models (general linear model, mixed linear model, FarmCPU, and BLINK). A total of 44 significant trait-SNP associations were identified, corresponding to 20 unique SNPs, as several SNPs (including those on chromosomes 6, 10, 15, and 17) were associated with multiple correlated traits. Linkage disequilibrium-based analysis showed candidate genes involved in phenylpropanoid metabolism and carbohydrate pathways. Predicted protein-altering variants, including non-synonymous and stop-gained mutations, were detected in key genes. Derived cleaved amplified polymorphic sequence markers developed near peak SNPs distinguished phenotypic differences between allelic groups, demonstrating their effectiveness for trait differentiation.

Conclusions

This study represents the first large-scale GWAS integrating image-based phenotyping for seed traits in perilla and provides candidate dCAPS markers with potential applicability to marker-assisted selection, pending validation in independent breeding populations. These findings offer valuable genetic resources and a practical framework for molecular breeding and crop improvement in perilla.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12870-026-09596-2.

Background

Perilla (Perilla frutescens) is a specialty oilseed crop widely cultivated in East Asia, particularly in Korea, China, and Japan [1, 2]. It is valued as a food crop and for its high content of bioactive compounds, including phenolics, flavonoids, and anthocyanins. These properties contribute to its nutritional and functional properties [3]. The accumulation and composition of these compounds are closely associated with seed characteristics and quality [4]. Therefore, seed-related traits are key targets for genetic studies and molecular breeding in perilla.

Among these traits, seed size is a critical agronomic factor affecting seed weight, yield components, and oil productivity. Larger seeds are associated with higher levels of fatty acids and bioactive compounds, contributing to improved nutritional quality and antioxidant capacity [5]. Seed size also influences germination and early seedling establishment, making it an important determinant of crop performance [6]. Seed coat color represents an important trait associated with phenolic composition and antioxidant activity [7]. Variation in pigmentation reflects differences in flavonoid and phenylpropanoid metabolism, which determine seed appearance and influence nutritional value and consumer preference [8].

Genetic studies of seed traits in perilla remain limited. Previous studies investigating seed size and seed coat color traits in perilla have primarily relied on relatively small populations or biparental mapping approaches, which restrict allelic diversity and mapping resolution [9–12]. Conventional phenotyping methods are often labor-intensive and prone to subjective bias, reducing the accuracy and reproducibility of trait evaluation. These limitations hinder the identification of robust marker–trait associations and restrict their applicability in breeding programs. For instance, seed size traits have traditionally been measured manually using rulers or calipers, and seed coat color has often been assessed through visual comparison or scoring, both of which are time-consuming and susceptible to inter- and intra-observer variation, particularly when applied to large germplasm panels [13]. Recent advances in image-based phenotyping provide an effective solution for high-throughput and objective trait measurement. Image-based methods enable precise and reproducible quantification of morphological and color-related traits while minimizing human bias [14, 15]. Image analysis tools such as ImageJ allow the extraction of detailed phenotypic parameters, including seed size components, for example, area, perimeter, and length, and color indices derived from RGB values [16, 17]. The integration of such high-resolution phenotyping with genome-wide association studies (GWAS) substantially improves the power and accuracy of detecting loci underlying complex traits [18].

Here, this study aimed to conduct a large-scale genome-wide association analysis by integrating high-throughput image-based phenotyping in a diverse panel of 493 Perilla frutescens accessions. This integrative approach improves the resolution and reliability of genetic analysis and bridges the gap between gene discovery and marker-assisted selection. This provides valuable resources for accelerating molecular breeding in perilla.

Methods

Plant materials and field conditions

A total of 493 accessions representing diverse genetic backgrounds were used (Table S1). The plant materials were cultivated at the experimental field of the Upland Crop Research Division, National Institute of Crop and Food Science, Rural Development Administration, Miryang, Republic of Korea. All the accessions were grown under standard cultivation practices recommended for perilla. More than five plants per accession were established with a spacing of 60 × 25 cm, with one plant maintained per hill. Seeds were sown in mid-June and harvested from September to November in 2023. Fertilizers were applied at the recommended rate of 4.0 kg N, 3.0 kg P, and 2.0 kg K per 1000 m² for perilla.

Image-based phenotyping of seed traits

Seed phenotyping was conducted using a high-throughput image-based phenotyping system that integrates image acquisition and analysis [19]. Approximately 40 seeds per accession were used as biological replicates for phenotypic measurements, and trait values were averaged across seeds within each accession to obtain a representative value per accession. Images were captured using a digital camera (SONY α6000) equipped with a macro lens (SEL30M35 mm F3.5, Sony, Tokyo, Japan) in a standardized imaging booth under controlled lighting conditions using a fixed camera distance to ensure consistent image acquisition. Image processing was performed using ImageJ software [20]. The workflow included scale calibration, image cropping, background removal, and color-threshold adjustment (Fig. 1A). Background regions were removed, and images were converted to binary format for pixel-based measurements. Individual seeds were segmented, and morphological traits including area, perimeter, maximum length, and minimum length were calculated. Seed coat color traits were quantified using RGB values extracted from the images. Prior to color analysis, a color reference chart was included in each image to calibrate and standardize RGB values across image sessions, ensuring consistency and comparability of color measurements. The R, G, and B values were averaged across seeds within each accession. The mean RGB value ((R + G + B)/3) was used as an indicator of brightness, where higher values correspond to lighter seed color.

Fig. 1.

Image-based phenotyping workflow and phenotypic variation of seed traits in Perilla frutescens. A Workflow for image-based phenotyping of perilla seeds using ImageJ. The process includes image acquisition under controlled conditions, grayscale conversion, thresholding, and binary image generation, followed by seed segmentation and labeling. Morphological traits such as area, perimeter, maximum length, and minimum length, and color traits (RGB values) were extracted from individual seeds. B Representative variation in seed size and seed coat color among perilla accessions. Selected germplasm samples illustrate the diversity in seed morphology and color intensity

DNA extraction and genotyping-by-sequencing (GBS)

Genomic DNA was isolated from young leaf tissues using the GenoAll^® Exgene™ Plant SV Kit (GenoAll Biotechnology, Seoul, Korea) following the manufacturer’s protocol. DNA quality and concentration were evaluated using a NanoDrop 3000 spectrophotometer (Thermo Fisher Scientific, USA). Samples were adjusted to a concentration of 20 ng µL⁻¹ prior to library preparation. Genotyping was carried out using the GBS method with two restriction enzyme strategies, namely, single digestion using ApeKI and double digestion using PstI-MspI. Library construction and sequencing followed established protocols [21–23]. Sequencing was conducted on the Illumina HiSeq X Ten platform. SNPs were detected and mapped to the reference genome of Perilla frutescens (cv. PF40) [24].

SNP calling, filtering, and imputation

Raw SNPs obtained from GBS were evaluated using a subset of 41 representative accessions selected from a single 96-plex sequencing plate to compare restriction enzyme strategies (ApeKI and PstI–MspI). SNPs in this subset were filtered using PLINK v1.9 with the following criteria: quality score ≥ 30, depth ≥ 5, missing rate ≤ 50%, bi-allelic loci, and presence of at least one homozygous alternative genotype (1/1). The SNP sets were used to assess SNP discovery and genotyping performance between enzyme combinations. SNP density was visualized using the SRplot platform [25].

For the full dataset for 493 germplasm, SNP filtering was performed using the following criteria: quality score ≥ 30, depth ≥ 5, bi-allelic SNPs, missing rate ≤ 30%, and minor allele frequency (MAF) ≥ 0.05. SNPs outside the assembled chromosomes were excluded. Missing genotypes were imputed using Beagle v5.4 with default parameters [26]. The final dataset was used for downstream analyses.

Population structure and linkage disequilibrium (LD) analysis

Population structure was assessed using principal component analysis (PCA) based on LD-pruned SNP markers. SNP pruning was performed using a sliding window approach (window size = 50 SNPs, step size = 5 SNPs, r² threshold = 0.2). Clustering analysis was conducted using the k-means algorithm based on the first three principal components. The optimal number of clusters was estimated using the elbow method. Population structure was further evaluated using STRUCTURE v2.3.4, with the number of clusters (K) ranging from 2 to 6. For each K, ten independent runs were performed, each consisting of a 50,000-iteration burn-in followed by 100,000 Markov Chain Monte Carlo (MCMC) iterations to ensure adequate convergence and reliable inference. LD analysis was conducted using PLINK v1.9 [27]. Pairwise LD (r²) was calculated for SNPs within a maximum distance of 1 Mb. LD decay was evaluated by grouping SNP pairs into distance bins and visualized using locally weighted regression curves in R (version 4.5.2).

Genome-wide association analysis (GWAS)

GWAS was conducted using the GAPIT package in R [28]. Four statistical models were applied, including the general linear model (GLM), mixed linear model (MLM), FarmCPU [29], and BLINK [30]. Population structure and kinship were incorporated to control for false-positive associations. The first three principal components were included as covariates in the GWAS models. GWAS was performed separately for each trait. SNPs detected by at least three of the four models were considered reliable associations. A Bonferroni-corrected significance threshold (P < 0.05/n) was applied.

Candidate gene identification and functional annotation

To compensate for the limited genome coverage of GBS, whole-genome resequencing data from 16 major Korean accessions were used to identify additional variants for candidate gene analysis [31]. Candidate genes were identified within LD blocks surrounding significant SNPs. SNP effects were predicted using SnpEff (version 5.2) [32]. Variants were classified based on their functional impact as high, moderate, low, and modifier. Gene annotation was performed using the perilla reference genome. Functional information was inferred through homology-based comparison with Arabidopsis thaliana (TAIR10). Protein sequences were compared using BLASTp (version 2.10.0), and candidate genes containing non-synonymous variants were prioritized.

Marker development and evaluation

Derived cleaved amplified polymorphic sequence (dCAPS) markers were developed to evaluate the phenotypic effects of alleles associated with seed size and seed coat color. The dCAPS markers were generated using dCAPS Finder 2.0 [33]. Appropriate restriction enzymes were selected to create or abolish restriction sites at the SNP positions by introducing mismatches in the primer sequences. PCR primers were designed using Primer3 [34] based on flanking sequences of target SNPs.

Statistical analysis

All statistical analyses were implemented using R (version 4.5.2). Descriptive statistics were calculated for all the phenotypic traits. Differences among groups were evaluated using analysis of variance (ANOVA). The results were visualized graphically using ggplot2 [35].

Results

Phenotypic variation of seed size and seed coat color

Analysis of seed size traits (area, perimeter, and length) and seed coat color traits (RGB, R, G, and B) in 493 perilla germplasm accessions revealed substantial phenotypic variation across all the measured traits (Fig. 1B; Table 1, and Figure S1). Seed area ranged from 1.47 to 8.52 mm² with a mean of 3.85 mm², while perimeter ranged from 4.47 to 10.92 mm (mean = 7.22 mm). Maximum and minimum lengths varied from 1.47 to 3.54 mm (mean = 2.36 mm) and from 1.28 to 3.07 mm (mean = 2.03 mm), respectively. For seed coat color traits, the RGB values ranged from 29 to 154 (mean = 72.44), with R, G, and B components showing similar variation across accessions. A wide range of phenotypic diversity was observed within the panel. This reflects substantial genetic variation for both seed size and color traits and supports its suitability for genome-wide association analysis.

Table 1.

Descriptive statistics of seed size and seed coat color traits in 493 Perilla frutescens accessions

Trait	Mean	Min	Max	Stdev1	Var2	Skew3	Kur4
Area (mm2)	3.85	1.47	8.52	1.20	1.44	0.48	1.31
Perimeter (mm)	7.22	4.47	10.92	1.18	1.40	−0.11	0.24
Max length (mm)	2.36	1.47	3.54	0.39	0.15	−0.13	−0.01
Min length (mm)	2.03	1.28	3.07	0.32	0.10	−0.05	0.54
RGB	72.44	29	154	30.19	911.58	0.77	−0.27
R	84.04	32	169	33.06	1092.71	0.51	−0.66
G	72.77	29	158	31.34	982.23	0.81	−0.21
B	60.42	26	138	26.77	716.37	1.02	0.23

¹Stdev standard deviation, ²Var variance, ³Skew skewness, ⁴Kurt kurtosis

Comparison of GBS enzyme strategies

To evaluate genotyping efficiency, a subset of 41 accessions was analyzed using two restriction enzyme strategies, namely, ApeKI and PstI–MspI. Sequencing reads, genome coverage, and SNP discovery were compared between the two approaches (Table S2). After applying standard filtering criteria, 40,882 and 25,079 SNPs were identified using ApeKI and PstI–MspI, respectively. ApeKI generated an average of 3,310,254 mapped reads per sample, with chromosome coverage of 3.5% and a mean sequencing depth of 9.2×. PstI–MspI produced a higher number of mapped reads (3,994,027 per sample) and greater sequencing depth (26.1×) but lower chromosome coverage (1.6%). Comparison of SNP sets indicated both shared and unique variants between the two methods (Fig. 2, Figure S2). This shows that each enzyme captures overlapping but distinct genomic regions. ApeKI generated a larger number of SNPs with broader genome coverage. Meanwhile, PstI–MspI produced higher sequencing depth but lower coverage. These results highlight the complementary strengths of the two enzyme strategies for SNP discovery. Both datasets were integrated, resulting in a final set of 34,848 SNPs for subsequent analyses (Table S3).

Fig. 2.

Genome-wide distribution of single-nucleotide polymorphisms (SNPs) generated by ApeKI and PstI–MspI in Perilla frutescens. SNP density across 20 chromosomes was visualized using genotyping-by-sequencing data generated with two restriction enzyme strategies, ApeKI (left) and PstI–MspI (right). SNPs were calculated within 1 Mb sliding windows. Color intensity represents the number of SNPs per window. ApeKI shows broader genome coverage with a higher number of SNPs. PstI–MspI exhibits more localized but deeper coverage, indicating complementary characteristics of the two enzyme systems

Population structure and trait variation

To characterize the genetic structure of the population, model-based clustering analysis was performed (Fig. 3). The number of clusters was evaluated using the elbow method, which suggested K = 3 or 4, and K = 3 was selected for subsequent analyses (Fig. 3A). STRUCTURE analysis supported this classification, with individuals partitioned into three subpopulations (Fig. 3B). PCA further validated the population structure, with the first two principal components explaining 30.43% and 19.43% of the total genetic variation, respectively (Fig. 3C). The PCA plot showed clear separation among groups. Accessions were colored according to K-means clustering based on the first three principal components. These clustering patterns were consistent with the STRUCTURE results, indicating clear genetic stratification within the panel.

Fig. 3.

Population structure and genetic diversity of Perilla frutescens accessions. A Determination of the optimal number of clusters using the elbow method based on within-cluster sum of squares. B Model-based population structure analysis showing the assignment of accessions at K = 2–4. C Principal component analysis (PCA) based on SNP markers, illustrating genetic differentiation among three subpopulations. D Boxplots of seed coat color traits (RGB, R, G, and B) across the identified clusters, showing significant phenotypic differences associated with population structure

To evaluate the relationship between genetic structure and phenotypic variation, seed traits were compared among subpopulations. Significant differences in seed coat color traits (RGB, R, G, and B) were observed among these three clusters (ANOVA, P < 0.05) (Fig. 3D). This indicates that population structure is associated with variation in seed coat color traits. The clear genetic stratification within the panel highlights the importance of incorporating population structure in the GWAS analysis.

LD decay pattern

Genome-wide LD decay was assessed using GBS-derived SNPs from 493 perilla accessions. LD (r²) decreased rapidly with increasing physical distance (Figure S3). Under a MAF ≥ 0.01 threshold, r² declined sharply within the first 200 kb and approached background levels at approximately 600–800 kb. Applying a stricter threshold (MAF ≥ 0.05) resulted in higher LD values across all distance classes. In both conditions, LD showed a rapid decline at short distances followed by a plateau at longer distances. The average r² decreased from approximately 0.08–0.10 to below 0.05 within approximately 200 kb. LD decayed to half of its maximum value at approximately 200 kb. This relatively rapid LD decay indicates that the panel is suitable for association mapping.

GWAS identifies loci for seed size and color traits

Genome-wide association analyses were conducted using BLINK, FarmCPU, MLM, and GLM. The results were evaluated using Manhattan and quantile–quantile (Q–Q) plots (Figs. 4A and 5A, Figure S4 and S5). SNPs detected by at least three of the four models were considered reliable and retained for further analysis. Based on this criterion, a total of 44 significant trait-SNP associations were identified, comprising 22 associations for seed size traits and 22 for seed coat color traits (Tables 2 and 3). The 44 associations corresponded to 20 unique SNPs (11 for seed size and 10 for seed coat color) because several SNPs were associated with more than one correlated trait. One SNP (Chr10_30859543) was shared between the two trait groups, being associated with seed area as well as all seed coat color components. Major loci were defined based on SNPs detected consistently across multiple models, including the BLINK model.

Fig. 4.

Genome-wide association analysis and linkage disequilibrium (LD)-based candidate region for Perilla frutescens seed size traits. A Manhattan plots showing genome-wide association results for seed size-related traits (area, perimeter, maximum length, and minimum length) using the BLINK model. The horizontal dashed line indicates the significance threshold (− log₁₀P = 6) based on Bonferroni correction using approximately 34,848 SNP markers. The highlighted region represents a major locus on chromosome 6 that is consistently associated with multiple traits. B Local association and LD structure surrounding the major locus on chromosome 6. The top panel shows SNP significance within the region, with the peak SNP (Chr06_28257681) highlighted in red and the linked SNP used for dCAPS marker development (Chr06_28248450) shown in blue. The middle panel displays the genomic positions of SNP markers within the candidate region. The bottom panel presents the LD heatmap based on pairwise r² values

Fig. 5.

Genome-wide association analysis and LD-based candidate region for Perilla frutescens seed coat color traits. A Manhattan plots showing genome-wide association results for seed coat color-related traits (RGB, R, G, and B) using the BLINK model. The horizontal dashed line indicates the significance threshold (− log₁₀P = 6) based on Bonferroni correction using approximately 34,848 SNP markers. The highlighted region represents a major locus on chromosome 10 that is consistently associated with multiple color traits. B Local association and LD structure surrounding the major locus on chromosome 10. The top panel shows SNP significance within the region, with the peak SNPs (Chr10_29579291 and Chr10_30859543) highlighted in red and the linked SNP used for dCAPS marker development (Chr10_30853195) shown in blue. The middle panel displays the genomic positions of SNP markers within the candidate region. The bottom panel presents the LD heatmap based on pairwise r² values

Table 2.

Genome-wide significant SNPs associated with seed size-related traits identified with GWAS in Perilla frutescens

Trait	Chr	SNP	Positions	Model (P-value)				MAF	LD block (bp)	No. of genes
				BLINK	FarmCPU	MLM	GLM
Area	3	Chr03_69858395	69,858,395	1.29E-06	2.71E-08	–	5.56E-11	0.054
	6	Chr06_28257681	28,257,681	6.69E-08	4.54E-10	–	1.44E-22	0.245
	10	Chr10_30853195	30,853,195	–	8.88E-08	3.96E-09	2.29E-07	0.381
	10	Chr10_30859543	30,859,543	5.94E-11	–	3.11E-09	8.88E-08	0.302
	11	Chr11_31051815	31,051,815	2.67E-07	1.20E-06	–	6.99E-09	0.393
	11	Chr11_45093681	45,093,681	2.57E-08	–	3.66E-09	5.52E-17	0.174
	15	Chr15_53218113	53,218,113	4.91E-18	–	2.56E-07	3.43E-17	0.205
	15	Chr15_53406122	53,406,122	–	1.94E-13	7.53E-08	1.62E-16	0.327
	17	Chr17_33185169	33,185,169	7.29E-08	8.13E-08	–	1.58E-17	0.103
Perimeter	6	Chr06_28257681	28,257,681	3.56E-08	1.24E-23	–	1.16E-26	0.245
	15	Chr15_53218113	53,218,113	3.94E-12	–	3.34E-07	1.57E-17	0.205
	15	Chr15_53406122	53,406,122	–	4.01E-12	1.17E-07	3.10E-16	0.327
	17	Chr17_33185169	33,185,169	9.41E-09	9.25E-09	–	1.43E-16	0.103
Max length	6	Chr06_28257681	28,257,681	2.07E-07	5.61E-20	–	5.46E-27	0.245
	11	Chr11_45093681	45,093,681	2.88E-09	–	5.33E-09	1.23E-16	0.174
	15	Chr15_53218113	53,218,113	1.98E-14	–	1.36.E-07	3.61E-18	0.205
	15	Chr15_53406122	53,406,122	–	3.36E-14	7.53E-08	1.25E-16	0.327
	17	Chr17_33185169	33,185,169	4.78E-11	4.00E-07	–	7.96E-18	0.103
Min length	2	Chr02_1707240	1,707,240	1.27E-08	1.05E-07	–	1.27E-14	0.13
	6	Chr06_28257435	28,257,435	3.26E-10	2.95E-23	–	2.42E-25	0.242
	11	Chr11_45093681	45,093,681	1.50E-09	–	3.84E-08	1.16E-13	0.174
	15	Chr15_53406122	53,406,122	–	2.11E-11	1.14E-06	9.97E-15	0.327
Seed size	6								28,018,819 − 28,538,548	8

Table 3.

Genome-wide significant SNPs associated with seed coat color-related traits identified with GWAS in Perilla frutescens

Trait	Chr	SNP	Positions	Model (P-value)				MAF	LD block (bp)	No. of genes
				BLINK	FarmCPU	MLM	GLM
RGB	3	Chr03_5139091	5,139,091	1.05E-10	9.36E-09	−	7.4E-13	0.365
	6	Chr06_1351191	1,351,191	5.86E-07	–	1.43E-06	1.93E-14	0.216
	8	Chr08_42926311	42,926,311	3.35E-08	5.91E-11	−	3.77E-13	0.471
	10	Chr10_4320952	4,320,952	–	4.93E-10	5.66E-08	1.70E-10	0.247
	10	Chr10_29579291	29,579,291	2.36E-15	2.28E-15	2.70E-23	3.18E-30	0.201
	10	Chr10_30859532	30,859,532	–	9.81E-47	8.12E-28	2.57E-25	0.334
	10	Chr10_30859543	30,859,543	9.03E-40	−	5.92E-29	1.22E-29	0.302
R	1	Chr01_72347042	72,347,042	5.64E-08	3.61E-10	−	3.73E-18	0.225
	3	Chr03_5139091	5,139,091	1.91E-13	1.04E-10	−	8.09E-14	0.365
	10	Chr10_4320952	4,320,952	1.30E-06	6.96E-08	1.18E-07	1.56E-09	0.247
	10	Chr10_29579291	29,579,291	2.71E-12	2.31E-13	1.21E-20	8.09E-29	0.201
	10	Chr10_30859543	30,859,543	6.48E-23	4.49E-23	4.55E-24	1.24E-24	0.302
G	8	Chr08_42926311	42,926,311	1.53E-08	2.59E-08	−	7.06E-13	0.471
	10	Chr10_4320952	4,320,952	−	8.26E-08	7.32E-08	1.26E-10	0.247
	10	Chr10_29575647	29,575,647	1.40E-10	−	1.03E-14	1.53E-14	0.382
	10	Chr10_29579291	29,579,291	7.45E-22	9.57E-11	1.28E-23	4.38E-30	0.201
	10	Chr10_30859543	30,859,543	9.02E-47	1.2E-36	3.63E-30	5.61E-31	0.302
B	6	Chr06_1351191	1,351,191	–	3.05E-07	1.33E-07	6.53E-15	0.216
	10	Chr10_4320952	4,320,952	1.81E-11	5.28E-09	4.18E-08	3.55E-11	0.247
	10	Chr10_29579291	29,579,291	1.92E-07	–	6.55E-25	5.58E-31	0.201
	10	Chr10_29963366	29,963,366	8.03E-09	–	3.49E-34	1.39E-30	0.331
	10	Chr10_30859543	30,859,543	1.56E-40	9.96E-63	4.10E-32	3.01E-33	0.302
Seed coat color	10								29,498,021–31,063,551	61

For seed size traits, significant loci associated with area, perimeter, maximum length, and minimum length were identified across chromosomes 3, 6, 10, 11, 15, and 17 (Table 2). A prominent locus on chromosome 6 (Chr6_28257681) was consistently associated with multiple traits, including area, perimeter, and length-related measurements (Fig. 4A). For seed coat color traits, significant associations were detected on chromosomes 3, 6, 8, and 10 (Table 3). A strong association signal on chromosome 10 (Chr10_30859543) was detected across multiple color components (RGB, R, G, and B) (Fig. 5A). This indicated a key genomic region controlling seed coat pigmentation.

Candidate genes underlying major loci

To identify candidate genes associated with significant SNPs, LD block analysis was performed for major loci detected in the GWAS. For seed size traits, the LD block on chromosome 6 (28.0–28.5 Mb) contained 8 genes and included the peak SNP (Chr6_28257681). This locus was associated with multiple traits, including area, perimeter, and length-related measurements (Fig. 4B). For seed coat color traits, the LD block on chromosome 10 (29.5–31.0 Mb) included 61 genes and had two peak SNPs (Chr10_30859543 and Chr10_29579291). Both peak SNPs were consistently associated with RGB, R, G, and B values (Fig. 5B).

In total, 69 genes were identified within the genomic regions spanning the LD blocks between flanking markers at the two major loci. SNP variants in these regions were annotated using SnpEff and classified into genomic categories, including upstream, coding sequence (CDS), intron, and downstream regions (Table S4). Functional annotation of these genes was performed by identifying homologous genes in Arabidopsis thaliana using BLASTp. Gene descriptions were assigned based on the best-matching Arabidopsis homologs. Among these, 57 CDS variants predicted to alter amino acid sequences were identified across 32 genes. This included missense, stop-gained, and splice region variants (Table 4).

Table 4.

Coding sequence variants predicted to alter protein sequences in 32 candidate genes

Chr	P. frutescens ID	A. thaliana ID	Position (bp)	Ref	Alt	Location	DNA/Protein level
Chr6	C2S51_014138	AT5G66980.1	28,244,961	C	G	Missense	c.116 C > G, p.Pro39Arg
	C2S51_014138	AT5G66980.1	28,245,259	G	A	Missense	c.331G > A, p.Gly111Arg
	C2S51_014138	AT5G66980.1	28,248,450	A	G	Missense	c.963 A > G, p.Ile321Met
	C2S51_014138	AT5G66980.1	28,248,545	C	G	Missense	c.1058 C > G, p.Thr353Arg
	C2S51_014138	AT5G66980.1	28,248,917	G	C	Missense	c.1228G > C, p.Val410Leu
	C2S51_014143	AT1G11790.1	28,495,532	G	A	Stop gained	c.1231 C > T, p.Arg411*
	C2S51_014144	AT1G62050.1	28,537,557	C	A	Missense	c.20 C > A, p.Thr7Lys
Chr10	C2S51_020324	AT5G23890.1	29,571,579	A	C	Missense	c.2126T > G, p.Leu709Arg
	C2S51_020327	AT5G23880.1	29,668,361	C	T	Missense	c.323 C > T, p.Ala108Val
	C2S51_020327	AT5G23880.1	29,669,925	A	C	Missense & splice region	c.908 A > C, p.Lys303Thr
	C2S51_020328	AT5G33406.1	29,709,667	A	G	Missense	c.91 A > G, p.Asn31Asp
	C2S51_020329	AT5G23860.2	29,748,318	G	A	Missense	c.916 C > T, p.Arg306Cys
	C2S51_020340	AT3G48890.1	29,959,040	C	T	Missense	c.665 C > T, p.Ala222Val
	C2S51_020341	AT5G52230.2	29,969,595	A	G	Missense	c.644 A > G, p.His215Arg
	C2S51_020343	AT1G72270.1	29,977,774	G	A	Missense	c.157G > A, p.Ala53Thr
	C2S51_020343	AT1G72270.1	29,977,906	A	G	Missense	c.289 A > G, p.Thr97Ala
	C2S51_020344	AT5G52190.1	29,993,524	A	C	Missense	c.491 A > C, p.Gln164Pro
	C2S51_020345	AT5G52180.1	30,005,973	A	G	Missense	c.851 A > G, p.Asn284Ser
	C2S51_020349	AT4G38180.1	30,122,329	T	A	Missense	c.2510 A > T, p.Gln837Leu
	C2S51_020350	AT3G17090.2	30,149,509	C	A	Missense	c.345G > T, p.Gln115His
	C2S51_020350	AT3G17090.2	30,150,417	C	T	Missense	c.70G > A, p.Glu24Lys
	C2S51_020353	AT5G62090.2	30,292,233	G	A	Missense	c.2005 C > T, p.Pro669Ser
	C2S51_020354	AT5G07220.1	30,390,328	C	G	Missense	c.718 C > G, p.Gln240Glu
	C2S51_020357	AT5G43050.1	30,539,668	G	C	Missense	c.222G > C, p.Arg74Ser
	C2S51_020357	AT5G43050.1	30,539,670	C	A	Missense	c.224 C > A, p.Ala75Glu
	C2S51_020358	AT5G52560.1	30,580,103	T	A	Missense	c.13 A > T, p.Thr5Ser
	C2S51_020359	AT3G28450.1	30,581,007	C	T	Missense	c.38 C > T, p.Thr13Ile
	C2S51_020359	AT3G28450.1	30,581,222	G	A	Missense	c.253G > A, p.Ala85Thr
	C2S51_020359	AT3G28450.1	30,581,286	T	C	Missense	c.317T > C, p.Leu106Pro
	C2S51_020360	AT5G59480.2	30,582,644	C	G	Missense	c.205G > C, p.Val69Leu
	C2S51_020360	AT5G59480.2	30,582,674	G	A	Missense	c.175 C > T, p.Pro59Ser
	C2S51_020360	AT5G59480.2	30,582,685	C	T	Missense	c.164G > A, p.Gly55Glu
	C2S51_020360	AT5G59480.2	30,582,716	G	A	Stop gained	c.133 C > T, p.Gln45*
	C2S51_020360	AT5G59480.2	30,582,790	T	C	Missense	c.59 A > G, p.Glu20Gly
	C2S51_020361	AT3G12950.1	30,587,383	A	G	Stop lost & splice region	c.589T > C, p.Ter197Glnext*?
	C2S51_020361	AT3G12950.1	30,587,789	C	T	Missense	c.520G > A, p.Val174Ile
	C2S51_020361	AT3G12950.1	30,588,126	C	T	Missense	c.265G > A, p.Asp89Asn
	C2S51_020361	AT3G12950.1	30,588,143	A	C	Missense	c.248T > G, p.Leu83Arg
	C2S51_020361	AT3G12950.1	30,588,153	G	A	Stop gained	c.238 C > T, p.Arg80*
	C2S51_020361	AT3G12950.1	30,588,162	C	G	Missense	c.229G > C, p.Glu77Gln
	C2S51_020361	AT3G12950.1	30,588,177	T	C	Missense	c.214 A > G, p.Ser72Gly
	C2S51_020361	AT3G12950.1	30,588,340	T	C	Missense	c.169 A > G, p.Asn57Asp
	C2S51_020362	AT2G02560.2	30,598,288	T	G	Missense	c.27T > G, p.Phe9Leu
	C2S51_020362	AT2G02560.2	30,598,392	T	G	Missense	c.131T > G, p.Ile44Ser
	C2S51_020362	AT2G02560.2	30,598,506	T	C	Missense	c.245T > C, p.Leu82Pro
	C2S51_020362	AT2G02560.2	30,598,696	A	G	Missense	c.341 A > G, p.Asn114Ser
	C2S51_020366	AT5G62170.1	30,853,195	C	G	Missense	c.405 C > G, p.Cys135Trp
	C2S51_020366	AT5G62170.1	30,855,011	T	G	Missense	c.1810T > G, p.Leu604Val
	C2S51_020369	AT5G51820.1	30,912,783	A	G	Missense	c.1189 A > G, p.Ser397Gly
	C2S51_020372	AT2G45510.1	30,975,059	A	G	Missense	c.113T > C, p.Phe38Ser
	C2S51_020373	AT5G66520.1	30,976,270	A	G	Missense	c.1432T > C, p.Tyr478His
	C2S51_020374	AT5G51740.1	30,981,222	C	T	Missense	c.253G > A, p.Val85Ile
	C2S51_020375	AT5G62190.1	30,986,253	C	A	Missense	c.478G > T, p.Asp160Tyr
	C2S51_020376	AT5G51660.1	30,993,839	C	G	Missense	c.246G > C, p.Met82Ile
	C2S51_020378	AT3G10710.1	31,019,666	G	A	Missense	c.139G > A, p.Val47Met
	C2S51_020379	AT5G51600.1	31,034,357	C	T	Stop gained	c.913 C > T, p.Arg305*
	C2S51_020380	AT4G12080.1	31,059,332	C	T	Missense	c.164 C > T, p.Thr55Met

Assessment of trait-associated markers for seed traits

Markers were developed as dCAPS from SNPs located within the same LD block as the peak SNPs identified for seed size (Chr6_28257681) and seed coat color (Chr10_29579291 and Chr10_30859543). Because the peak SNP positions did not contain restriction enzyme recognition sites suitable for dCAPS design, we instead selected proxy SNPs that were located within the same LD block, were in linkage disequilibrium with the corresponding peak SNP, and permitted dCAPS primer and restriction-site design (Table S5). For seed size, Chr6_28248450 was selected as a proxy for the peak SNP Chr6_28257681, located 9.2 kb away (r² = 0.88, strong LD). For seed coat color, Chr10_30853195 was selected as a proxy for the peak color SNP Chr10_30859543, located 6.3 kb away (r² = 0.70). Although Chr10_30853195 is listed in Table 2 as a seed area–associated SNP, it lies within the chromosome 10 LD block (29.5–31.0 Mb) that also harbors Chr10_30859543, the single SNP jointly associated with seed area and all seed coat color components (Tables 2 and 3). Its linkage with the peak color SNP therefore makes it a suitable linked marker for tagging seed coat color variation at this locus.

These markers were applied to the accessions, which were grouped according to their marker genotypes. Significant differences in seed size and seed coat color were observed among the genotype groups (Fig. 6). For seed area at Chr6_28248450, accessions carrying the AA allele showed a higher mean value of 4.31 than those with the GG allele, which had a mean of 2.43. This indicated that the AA allele is associated with increased seed size. For RGB values at Chr10_30853195, accessions carrying the CC allele had a higher mean value of 78.96 compared to those with the GG allele, which had a mean of 69.37. This suggests that the CC allele is associated with a brighter seed coat color. These differences were statistically significant with P < 0.01. Allelic grouping based on each marker differentiated accessions for both area and RGB. This supports the potential utility of these markers for marker-assisted selection across diverse genetic backgrounds.

Fig. 6.

Allelic effects of dCAPS markers on Perilla frutescens seed size and color traits. Boxplots showing phenotypic differences between allelic groups of dCAPS markers across 493 accessions. A The marker on chromosome 6 (Chr06_28248450) is associated with seed area (left), and B the marker on chromosome 10 (Chr10_30853195) is associated with seed coat color (RGB) (right). Accessions were grouped based on genotype, and statistical significance between allelic groups was evaluated using t-tests. Asterisks indicate significant differences (**P < 0.01, ***P < 0.001)

Discussion

In this study, the genetic architecture of seed size and seed coat color in Perilla frutescens was analyzed by combining high-throughput image-based phenotyping with large-scale genome-wide association analysis. Compared with previous studies limited by small populations and conventional phenotyping, the current study using 493 diverse accessions and objective measurements improved phenotypic accuracy and detection power. The identification of loci consistent across multiple GWAS models supports the robustness of the marker–trait associations.

The reliability of the GWAS results is supported by the observed population structure and LD patterns. Three genetic groups were identified, and incorporating population structure into GWAS models reduced spurious associations [36]. LD decayed rapidly with increasing physical distance [37], supporting the suitability of the panel for association mapping. Nonetheless, some candidate intervals, such as the chromosome 10 region (29.5–31.0 Mb), remained relatively large, and mapping resolution was not uniform across loci.

The genetic architecture of seed size traits suggests a central role for metabolic pathways in seed development. A stop-gained mutation was identified in C2S51_014143, homologous to Arabidopsis thaliana AT1G11790 (ADT1), a key enzyme in phenylalanine biosynthesis [38]. Disruption of ADT genes has been associated with altered carbon allocation and seed development, including reduced kernel size in maize [39]. This premature stop codon suggests that changes in metabolic flux contribute to seed size variation in perilla.

Seed coat color variation was associated with genes involved in phenolic metabolism and pigment biosynthesis. The gene C2S51_020363, homologous to AT5G51890, encodes a class III peroxidase involved in phenolic polymerization and seed coat pigmentation [40]. C2S51_020369, homologous to AT5G51820 (PMG1), encodes plastidic phosphoglucomutase, which may influence pigment accumulation through carbohydrate-mediated regulation of flavonoid biosynthesis [41, 42]. Both phenolic modification and upstream metabolic processes regulate perilla seed coat color. Although candidate genes were identified based on LD structure and functional annotation, further experimental validation is required to confirm their biological roles.

Integrating image-based phenotyping with GWAS improves the analysis of complex traits in perilla. Manual measurement of seed size using rulers or calipers, and visual scoring of seed coat color, are inherently limited by measurement precision and observer subjectivity, which can obscure subtle but genetically meaningful phenotypic variation, particularly across a genetically diverse panel of hundreds of accessions. Compared with manual measurements, image-based methods provide more precise and high-throughput phenotypic data. This improves the accuracy and reproducibility of genetic analyses [14, 15]. Using RGB-based color quantification and detailed morphological traits enabled detection of robust associations not easily captured by conventional methods. This approach provides a scalable framework for genetic analysis in perilla and other underutilized crops [18].

The markers developed in this study differentiated allelic groups within the discovery panel, indicating their potential as candidate tools for marker-assisted selection [43]. Despite these findings, several limitations of the present study should be considered. Because marker testing was conducted within the same association panel used for GWAS discovery rather than in an independent population, this result demonstrates allelic differentiation within the discovery panel rather than independent validation or confirmed functional utility [44]. Phenotypic evaluation was conducted in a single environment and growing season, and further testing across multiple environments and in independent, genetically diverse breeding populations will be required before these markers can be considered validated tools for routine use in breeding programs [45]. In addition, genotyping was based on GBS, which provides limited genome coverage [21]. Although whole-genome resequencing data from a subset of accessions were used to supplement candidate gene analysis, some causal variants outside the captured regions may not have been detected [21]. Furthermore, candidate gene prioritization relied primarily on homology-based annotation with Arabidopsis thaliana and predicted coding-sequence variant effects [46]. This approach identifies biologically plausible candidates, but additional evidence such as haplotype-based association analysis, expression data, and functional validation will be required to confirm their causal role [47, 48]. Future work will focus on validating these markers in independent, multi-environment breeding populations and on functional characterization of the candidate genes through haplotype and expression analysis.

Although there are some limitations, this study provides a useful foundation for developing markers in perilla, a crop for which there are not many resources available for genomic and molecular breeding. The large population size and objective image-based phenotyping enabled reliable trait quantification, and the consistency of marker–trait associations across multiple GWAS models strengthens confidence in the identified loci. Together with the trait-associated markers evaluated here, these results offer a valuable starting point for marker-assisted selection and further genetic dissection of the target trait in perilla.

Conclusions

In this study, the genetic architecture of seed size and seed coat color in Perilla frutescens was elucidated by combining high-throughput image-based phenotyping with large-scale GWAS in a diverse panel of 493 accessions. Stable loci were identified across multiple analytical models. Candidate gene analysis showed key pathways related to phenylpropanoid metabolism and carbohydrate regulation. The dCAPS markers developed in this study indicated clear allelic differentiation for seed size and seed coat color within the discovery panel, indicating their potential utility for marker-assisted selection. These findings provide candidate markers with potential utility for marker-assisted selection, pending validation in independent populations. This highlights the effectiveness of combining image-based phenotyping with genomic analysis to accelerate crop improvement in perilla.

Abbreviations

GWAS

Genome-wide association study

SNP

Single-nucleotide polymorphism

dCAPS

derived cleaved amplified polymorphic sequence

GBS

Genotyping-by-sequencing

MAF

Minor allele frequency

LD

Linkage disequilibrium

PCA

Principal component analysis

ANOVA

Analysis of variance

CDS

Coding sequence

Authors’ contributions

Y.J.A. conceived and coordinated the study, performed formal analysis and visualization, and wrote the original draft. E.Y., J-I.K. and S.W.K. provided resources and reviewed the manuscript. S-K.P., Y-H.Y., G.P., M.H.L., S-H.K. reviewed the manuscript. E.L. acquired funding, managed the project, conceptualized the study, developed the methodology, conducted the investigation.

Funding

This research was supported by the Cooperative Research Program for Agriculture Science and Technology Development (Project PJ017387072026), Rural Development Administration, Republic of Korea. It was also supported by the 2026 collaborative research program between university and the Rural Development Administration, Republic of Korea.

Data availability

The datasets used for identifying candidate variants (whole-genome resequencing of 16 Korean Perilla cultivars) are available in the NCBI BioProject database (ID: PRJNA1276257). Any datasets generated or analysed during the current investigation are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable. This study did not involve humans or experimental animals.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.