Development and deployment of a public mid‐density genotyping panel for crimson clover (Trifolium incarnatum L.) based on whole‐genome re‐sequencing of diverse germplasm

Abstract

Crimson clover (Trifolium incarnatum L.) is an obligately outcrossing, cool‐season annual legume valued for forage and cover cropping, yet genomic resources to support systematic improvement are limited. We performed the first whole‐genome re‐sequencing of global crimson clover germplasm to (i) develop a public mid‐density enrichment capture panel for breeding applications and (ii) characterized genetic diversity among the sampled genotypes to validate the representativeness of the germplasm for panel design. From 45 crimson clover accessions, we sequenced 194 genotypes: 45 at ∼50X, generating 5.84 million variants and 149 at ∼2.54X, yielding 17.05 million variants. After filtering, we retained 542,790 high‐confidence single‐nucleotide polymorphisms (SNPs) from the high‐coverage dataset and ∼2.4 million from the low‐pass cohort, respectively. From these resources, we designed a 28,867‐SNP Twist hybrid‐capture panel enriched for genic regions and evenly distributed across seven chromosomes. This panel is being deployed within Auburn University's crimson clover breeding program to support population improvement and cultivar development. Genetic diversity analyses (principal component analysis, ADMIXTURE) revealed compact clustering of cultivars, broader dispersion of wild and uncertain‐status accessions, and low overall differentiation (fixation index [FST] = 0.0105) with excess heterozygosity (inbreeding coefficient [FIS] = −0.0592), consistent with a high level of outcrossing. The resulting genomic resources provide a reproducible, mid‐density genotyping platform for trait discovery, predictive breeding, and diversity monitoring. Together, these advances bring crimson clover genomic resources on par with other legumes such as soybean (Glycine max L. Merr.) and alfalfa (Medicago sativa L.), establishing a robust foundation for genomics‐assisted improvement of this key cover and forage crop in US sustainable agriculture.

Core Ideas

• Whole‐genome re‐sequencing of 194 crimson clover accessions revealed >21 million variants.

• High‐confidence single‐nucleotide polymorphism (SNP) catalogs from 50X and 2X data enable cost‐effective genotyping.

• Genetic diversity is weakly structured, with cultivars clustering narrowly by origin.

• A 28,867 SNP enrichment panel delivers uniform genome coverage and >75% genic content.

• These genomic tools accelerate GWAS, genomic selection, and breeding innovation.

Plain Language Summary

Crimson clover is among the most widely planted legume forage and cover crops in the United States, where it protects farmland, builds soil fertility, and reduces dependence on synthetic fertilizers. Yet until now, no genome‐wide genetic tools existed to accelerate its improvement, leaving breeders without the resources available for crops like soybean and alfalfa. This study addresses that gap. We sequenced 194 globally sourced plants at high and low depths to capture millions of genetic differences, then developed and deployed the first publicly available genotyping panel for crimson clover breeding, measuring nearly 30,000 genetic markers across the genome. Validation on over 1000 breeding plants confirmed that the panel reliably detects more than 50,000 useful genetic differences. This tool is commercially manufactured and available to breeding programs worldwide, providing a foundation for improved varieties that support soil health, food production, and sustainable agriculture.

Abbreviations

AFLP

amplified fragment length polymorphism

AU

auburn University

AUHPC

auburn University High‐Performance Computing; FIS, inbreeding coefficient; FST, fixation index; GWAS, genomic wide association studies

JGI

joint Genome Institute

LD

linkage disequilibrium

MAF

minor allele frequency; NGS, next‐generation sequencing

PCA

principal component analysis; PCR, olymerase chain reaction

PYT

preliminary yield trial; REF, reference allele

SNP

single‐nucleotide polymorphism; SPAD, soil plant analysis development

SPT

spaced plant trial; SSR, simple sequence repeats

UF

University of Florida

WGR

whole‐genome re‐sequencing

1. INTRODUCTION

Crimson clover is a cool‐season, obligately outcrossing legume valued worldwide as both a forage and cover crop due to its high biomass production, strong nitrogen fixation capacity, and contributions to soil fertility and ecosystem resilience (Barnes et al., 1995; Moore et al., 2020). By reducing reliance on synthetic fertilizers and enhancing nutrient cycling, crimson clover plays a vital role in sustainable intensification across diverse cropping systems. Recent surveys underscore its growing agricultural importance, such as the 2022–2023 National Cover Crop Survey, which ranks crimson clover first among legumes and fourth among all cover‐crop species planted in US fields, behind only cereal rye (Secale cereale L.), radish (Raphanus sativus L.), and winter wheat (Triticum aestivum L.), and with many growers reporting more than a decade of continuous use, reflecting both agronomic reliability and long term soil benefits (Conservation Technology Information Center (CTIC) et al., 2023). Today, crimson clover seed production is approximately 2.6–4.4 million kg of seed annually, valued at $3.9–$6.2 million in farm‐gate sales. Production is concentrated in Oregon's Willamette Valley, which now supplies nearly all US crimson clover seed. Between 2018 and 2020, harvested area in Oregon ranged from 2315 to 3650 ha, with average yields of approximately 1120 kg ha^{− 1} (Oregon State University Extension Service, 2021).

Despite its agronomic and ecological importance, crimson clover has historically lagged far behind other forage and grain legumes in genomic resource development. While model and crop legumes such as soybean, alfalfa, and red clover (Trifolium pratense L.) now benefit from chromosome‐scale reference genomes, dense single‐nucleotide polymorphism (SNP) arrays, and large‐scale diversity studies that have transformed molecular breeding (Bickhart et al., 2022; Varshney et al., 2021), comparable genomic resources for crimson clover remain limited. Within the genus Trifolium (T.), most genomic studies have focused on red and white clover (Trifolium repens L.). Red clover now possesses multiple chromosome‐level assemblies supporting genomic wide association studies (GWAS), quantitative trait loci  mapping, and genomic selection for key forage agronomic traits (Bickhart et al., 2022; Jones et al., 2020; Yan et al., 2022). While white clover, an allotetraploid, has similarly high‐contiguity genome assemblies that reveal recombination landscapes and inform polyploid prediction (Kuo et al., 2024). In contrast, our only knowledge about crimson clover includes a C‐value genome size estimate (∼652 Mb; 2n = 14) and an early amplified fragment length polymorphism (AFLP)‐based diversity surveys (Ellison et al., 2006; Steiner et al., 1998; Vižintin et al., 2006). The absence of whole genome assemblies and genome‐wide markers is constraining breeding progress in crimson clover.

Reduced‐representation methods such as simple sequence repeats (SSR) or genotyping‐by‐sequencing (GBS) have provided only partial insight, capturing only a small fraction of the genome and producing marker sets that are often inconsistent across studies. Because GBS relies on restriction enzyme digestion, it often suffers from missing data, locus dropout, and batch effects that limiting reproducibility and downstream application (Elshire et al., 2011; Poland & Rife, 2012). In contrast, whole‐genome re‐sequencing (WGR) offers comprehensive variant discovery across coding and noncoding regions, enabling dense SNP catalogs, accurate estimates of linkage disequilibrium (LD), and the foundation for genomic tools that can accelerate breeding (Tester & Langridge, 2010; Varshney et al., 2014). Expanding such genomic resources for crimson clover is therefore a critical need.

While WGR remains too costly for routine breeding applications (genomic selection, GWAS profiling), it provides the discovery layer for designing efficient, reduced‐representation genotyping tools. One such approach is hybrid‐capture enrichment panels, probe‐based genotyping platforms that selectively sequence thousands of predefined genomic regions. Hybrid capture employs short synthetic oligonucleotide “baits” that hybridize to target loci, enabling reproducible, genome‐wide SNP recovery at moderate cost (Lemmon et al., 2012; Twist Bioscience, 2023). Compared with GBS, enrichment panels offer consistent marker sets, reduced missing data, and improved cross‐study comparability, making them ideally suited for genomic selection, GWAS, and germplasm characterization.

To fill this gap, we conducted the first WGR study of crimson clover germplasm. We applied a two‐tiered re‐sequencing strategy with a core set sequenced at high (∼50X) coverage and a broader set at low (∼2X) coverage; this approach balances coverage, resolution, and cost. The specific objectives were to (i) generate a comprehensive SNP resource to support genomic tool development; (ii) develop a publicly available, mid‐density genome‐wide hybrid‐capture enrichment panel (∼28K SNPs) optimized for crimson clover breeding applications; and (iii) characterize genetic diversity across global panel to validate the representativeness of the germplasm used for panel design. Together, these efforts establish the first genome‐wide variant resource for crimson clover and provide a foundation for bridging variant discovery and applied breeding, bringing this agronomically important legume into the genomics era.

Core Ideas

• Whole‐genome re‐sequencing of 194 crimson clover accessions revealed >21 million variants.

• High‐confidence single‐nucleotide polymorphism (SNP) catalogs from 50X and 2X data enable cost‐effective genotyping.

• Genetic diversity is weakly structured, with cultivars clustering narrowly by origin.

• A 28,867 SNP enrichment panel delivers uniform genome coverage and >75% genic content.

• These genomic tools accelerate GWAS, genomic selection, and breeding innovation.

2. MATERIALS AND METHODS

2.1. Plant material and sample collection

A total of 45 distinct crimson clover accessions were collected for this study, representing a mixture of synthetic populations and maternal half‐sib families. Specifically, 40 accessions were obtained from the USDA National Plant Germplasm System (NPGS), encompassing plant introductions (PIs) from diverse geographic origins. NPGS accessions generally represent small synthetic populations of open‐pollinated individuals with unknown effective population sizes, reflecting natural diversity. In contrast, accessions provided by Auburn University (AU) (two accessions) and the University of Florida (UF) (two accessions) were derived from maternal half‐sib families, each representing progeny of open‐pollinated plants selected within breeding nurseries. The AU accessions consisted of one polycross family (AUCC14) originating from a cross between the cultivars AU Robin and AU Reseeding. The UF accessions included two half‐sib families (UFCC17 and UFEarlyWhite). In addition, the commercial cultivar Dixie (one accession) was included as a replicated check. Seeds were sown in October 2022, with 12 seeds per family in 72‐cell trays. After 6 weeks, seedlings were transplanted to the E.V. Smith Research Station—Plant Breeding Unit in Tallassee, AL (32.5117° N, 85.8925° W). Plants were spaced 18 in. apart in a nursery layout following an augmented design, with the cultivar Dixie included as a check in each row, totaling 350 plants at establishment. Following a winter freeze event, 194 plants survived and were retained for tissue sampling, consisting of 154 PI plants, 18 AU plants, five UF plants, and 17 Dixie check plants. Young leaf tissue was collected in early 2023 during late vegetative and early reproductive stages, flash‐frozen in liquid nitrogen, and stored at −80°C until DNA extraction.

2.2. DNA extraction

Genomic DNA was extracted from lyophilized tissues using a modified cetyltrimethylammonium bromide protocol optimized for leguminous species (Doyle, 1990). This protocol incorporated additional steps to remove polysaccharides and secondary metabolites. DNA quantity and quality were assessed using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific) and Qubit 4 Fluorometer (Invitrogen), following standard manufacturer protocols.

2.3. Two‐tiered sequencing strategy

The 194 individual plants (unique genotypes) were sequenced as part of a two‐tiered WGR strategy (Figure 1). For each of the 45 accessions, up to 10 plants were transplanted to the field to capture within‐accession diversity. Because crimson clover is genetically heterogeneous, multiple individuals per accession were evaluated to capture within‐accession genetic variation. Three phenotypic traits were scored: plant height, vigor (visual score, 3 < 5 < 7), and soil plant analysis development (SPAD) (a proxy for chlorophyll and nitrogen content). A linear mixed model was fitted for each trait to account for field blocking effects, and best linear unbiased predictors were obtained. A Smith‐Hazel selection index (Hazel, 1943; Smith, 1936) was constructed to combine these traits into a single composite score. The index used economic weights of 0.5, 0.3, and 0.2 for height, vigor, and SPAD, respectively, emphasizing biomass and plant robustness. Because pedigree or genomic relationships were not yet available, the index was applied at the phenotypic level to identify the best individuals within each family. For each accession or family, one top‐ranked plant based on the index was selected for high‐coverage sequencing (∼50X), while the remaining individuals were sequenced at low coverage (∼2X) to maximize population representation. In total, 45 plants were sequenced at high coverage (50X) and 149 at low coverage (2X), forming a discovery‐plus‐diversity dataset.

FIGURE 1.

Overview of Auburn crimson clover breeding program genomics workflow and resources. (a) Joint Genome Institute (JGI) NGS (next‐generation sequencing) service and data storage. (b) Twist DNA synthesis vendor. (c) Application of genomic data within the breeding pipeline to support marker discovery.

2.4. Library preparation and sequencing

Genomic DNA from all samples was randomly sheared to ∼500 bp and used to prepare Illumina paired‐end libraries. Sequencing was performed on the Illumina NovaSeq 6000 and Illumina NovaSeq X, generating 150 bp paired‐end reads. Quality control and adapter trimming were applied to raw reads prior to alignment to the T. pratense reference genome (GCF_020283565.1_ARS_RC_1.1) using BWA‐MEM v07.17 (H. Li & Durbin, 2009). Variant calling was conducted using SAMtools v1.11 (H. Li et al., 2009) and bcftools v1.11 (Danecek et al., 2021) mpileup. All data, including FASTQs files (sequence and quality score format), binary alignment/map files, and variant call files (VCFs), are to be publicly accessible via the Department of Energy (DOE) Joint Genome Institute (JGI) Genome Portal under Proposal ID 509088.

2.5. Sequence quality control, read alignment, and variant discovery

Raw Illumina NovaSeq 6000 and Illumina NovaSeq X 2 × 151 bp reads were initially processed using DOE JGI following Resequencing SOP 1099.3 (DOE Joint Genome Institute, 2023). Read filtering and quality improvement were performed using the BBTools v39.03 RQCFilter2 workflow (Bushnell, 2014). This pipeline includes optical and polymerase chain reaction (PCR) duplicate removal (Clumpify), adapter trimming, removal of terminal poly‐G sequencing artifacts, right‐end quality trimming below Q6, and filtering of reads containing ambiguous bases or failing minimum‐length requirements. The workflow also screens against a comprehensive contamination database (PhiX, microbial, organellar, and animal references), though no contaminant reads were detected across libraries. After filtering, the retained high‐quality reads represented the nonredundant, informative sequence content of each library. High‐quality reads were aligned using BWA‐MEM v0.7.17 with default setting to the T. pratense reference genome (GCF_020283565.1 ARS_RC_1.1; Bickhart et al., 2022). This reference was selected as the best available same‐section (basic chromosome number [x = 7]), chromosome‐scale Trifolium reference for Trifolium incarnatum (Ellison et al., 2006; Vižintin et al., 2006). Alignment sorting, indexing, and processing were performed using SAMtools v1.11 and variant calling was conducted on the Auburn University High‐Performance Computing (AUHPC) system using bcftools v1.11 mpileup followed by bcftools call to generate per‐sample VCFs. Individual VCFs were subsequently merged through joint variant calling to produce a unified multi‐sample dataset for each coverage level. Variants were filtered with bcftools filter using quality and depth thresholds (QUAL > 30 and DP > 10), retaining only biallelic SNPs. Additional filters included a minor allele frequency (MAF) ≥ 0.05 and a missing genotype rate ≤ 0.10, ensuring that retained sites were genotyped in at least 90% of individuals. The high‐coverage (50X) and low‐coverage (2X) datasets were processed independently through similar variant‐discovery and filtering pipelines implemented with bcftools v1.11 and PLINK 1.9 (Chang et al., 2015). After variant normalization with bcftools v1.11 norm, haplotype phasing and genotype imputation were performed using Beagle v5.4 (Browning et al., 2018, 2021). Complete filtering parameters and workflow details are provided in Table 1.

TABLE 1.

Single‐nucleotide polymorphism (SNP) filtering strategy for the 50X whole‐genome re‐sequencing (WGR) dataset of 45 Trifolium incarnatum genotypes, yielding 542,790 high‐confidence SNPs for downstream analyses.

Filtering step	Criteria applied	Remaining SNPs	Percentage (%)
Total variants identified	Initial variant calling (50X coverage)	5,843,860	100
SNPs identified	SNP subset from total variants	4,958,826	84.86
30% missing data filter applied	Removal of SNPs with >30% missing data	1,698,625	29.07
Minor allele frequency (MAF < 0.05)	MAF threshold set at 0.05	542,790	9.29
Beagle 5.4 imputation applied	Imputation for missing genotypes	542,790	9.29
Final high‐confidence SNPs	Quality‐controlled dataset	542,790	9.29

2.6. Targeted enrichment panel design

For the reason explain above, our workflow was anchored to the red clover (ARS_RC_1.1) reference assembly (RefSeq GCF_020283565.1). Coordinate conventions were harmonized across file formats by treating VCF positions as 1‐based and BED intervals as 0‐based end‐exclusive, and by enforcing exact chromosome naming consistency with the FASTA index (NC_060059.1–NC_060065.1).

Gene models from the ARS_RC_1.1 GFF annotation were merged to define a comprehensive “genic union,” which included entire gene bodies and their associated exon, coding sequence (CDS), and untranslated region (UTR) features. Candidate SNP discovery integrated three independent variant sources: a high‐coverage (∼50X) JGI panel, a low‐coverage (∼2X) JGI panel, and a panel of 6803 genome‐wide SNPs selected because they intersected variants discovered in a separate sample of crimson clover breeding germplasm (data not shown) genotyped using Khufu in 2023. Khufu is a low‐pass whole‐genome sequencing and genotyping platform developed at HudsonAlpha Institute for Biotechnology that maps short reads at ∼1–3X coverage and applies filtering algorithms for high‐confidence SNP identification followed by panel‐free haplotype imputation (https://www.hudsonalpha.org/khufudata/).The Khufu SNPs were included to supplement the JGI‐derived variant catalog with markers validated as polymorphic in applied breeding populations. To avoid redundancy, Khufu SNPs that overlapped with existing JGI variant positions were excluded. In the end, 6803 additional markers represented genomic regions not already captured by the JGI discovery panels were added. These markers were filtered to retain only sites with ≤80% missing data and allele frequencies (AFs) ranging from 0.05 to 0.95. All VCF files were normalized against the same reference FASTA to ensure consistency. This process involved left‐aligning and decomposing multiallelic sites, standardizing reference (REF)/advanced line trials (ALT) allele representations, and then compressing and indexing the outputs prior to joint processing. All analyses were conducted on the AUHPC. The pipeline utilized version‐controlled AUHPC modules, including bcftools v1.20, htslib v1.20 (Bonfield et al., 2021), samtools v1.20, bedtools v2.31.1 (Quinlan & Hall, 2010), and plink2/2.00‐alpha (Chang, 2023). Software versions are reported as loaded on AUHPC between July and August 2025 to ensure computational reproducibility and transparency.

Each sequencing strategy produced distinct dataset characteristics and post‐next generation sequencing (NGS) specifications requiring customized workflow optimization. In this phase, variant calls from the high‐coverage (50X) and low‐coverage (2X) JGI datasets were anchored to the red clover reference genome (RefSeq GCF_020283565.1). The two variant call sets were merged with bcftools v1.20 merge to generate a comprehensive variant catalog for probe design. Site‐level quality control was implemented with bcftools 1.20 view and vcftools v0.1.16 (Danecek et al., 2011), retaining only biallelic SNPs with QUAL ≥ 30, MAF ≥ 0.03, and per‐site missingness ≤ 0.10. The slightly relaxed MAF threshold followed predictive‐breeding panel‐design principles to include moderately rare alleles that enhance marker diversity and genome coverage. Strand‐ambiguous polymorphisms (A/T and C/G) were identified and excluded using a custom awk/grep filter, while LD pruning was carried out in plink 1.9 with a 50‐kb window, 5‐kb step, and r ² ≤ 0.2 (validated at r ² ≤ 0.1) to ensure independence among retained variants.

Each variant was annotated as genic or nongenic via bedtools (Quinlan & Hall, 2010) intersect, comparing variant coordinates against a merged genic union GFF3 annotation derived from the ARS_RC_1.1 reference. Overlapping records were resolved deterministically using a short Python 3.10 script that prioritized genic status and higher MAF values. Capture targets were then defined by expanding each site ±60 bp with bedtools slop, merging overlaps with bedtools merge, and converting outputs into BED3 format. To enforce chromosome‐proportional coverage, a bin‐center heuristic implemented in Python (pandas + numpy; Harris et al., 2020) partitioned each chromosome into equal‐length bins and selected the variant nearest the bin midpoint, redistributing gaps iteratively to maintain balanced genomic coverage. The optimized JGI‐derived set was unified with a curated Khufu‐filtered variant pool using bcftools isec and bedtools merge to collapse overlapping or adjacent intervals. A final “top‐up” procedure scripted in Python favored genic features and ensured uniform spacing across the genome. Chromosome‐level distribution plots were generated on AUHPC using Python 3.10, CairoSVG (Murray et al., 2024) ≥2.7, and Pillow v10.4.0 (Murray et al., 2024) ≥10 within a minimal virtual environment. Visualization scripts parsed FASTA indices and BED3 coordinates to produce high‐resolution portable network graphics  labeled with both “Chr01–Chr07” and reference accession identifiers.

In silico probe validation was conducted using Bowtie2 v2.4.2 (Langmead & Salzberg, 2012). Each ±60 bp window was converted to a 121‐bp synthetic probe sequence with bedtools getfasta, and the reference genome was indexed via bowtie2‐build and synchronized with samtools faidx. End‐to‐end alignments were executed under the –very‐sensitive preset with one seed mismatch (‐N 1) and a reporting limit of two placements per probe (‐k 2). A secondary stress test using local alignments (–very‐sensitive‐local, ‐L 20, ‐N 1, ‐k 5) identified potential partial or clipped matches. Probes mapping uniquely under both alignment modes were retained for synthesis. Following in silico validation, the final probe manifest was submitted to Twist Bioscience (https://www.twistbioscience.com/products/ngs/custom‐panels.), where internal quality control confirmed probe balance, hybridization potential, and manufacturing feasibility. Twist's verification ensures compatibility with capture chemistry and production, providing an additional layer of validation beyond computational assessment.

2.7. Twist hybrid capture enrichment panel verification using imputed HapMap genotypes

To evaluate the genotyping performance of the Twist hybrid‐capture enrichment panel, we applied it to 1039 crimson clover breeding samples in 2024 and 2025 (genotyping scheme shown in Figure 1). Detailed analysis of the resulting dataset will be published elsewhere. Our objective here is to demonstrate the effectiveness of the enrichment‐panel in terms of genotyping completeness and AF spectrum using breeding‐relevant samples. HapMap genotype datasets were generated as part of this workflow. The dataset was derived from enriching the sequenced reads for target loci using oligonucleotide probes (as defined in the panel BED files). Khufu variant calling and imputation was subsequently applied. Genotyping performance was evaluated using this dataset. HapMap files were imported into R (R Core team, 2025) (data.table; Barrett et al., 2026), and genotype strings were converted to additive allele dosages (zero, one, or two copies of the alternate allele) using marker‐specific allele definitions provided in the HapMap alleles field. Non‐informative or invalid genotype entries (e.g., missing calls, ambiguous bases, malformed strings, or alleles inconsistent with expected REF/ALT states) were set to NA prior to transformation. Dosage conversion was parallelized across samples (parallel::mclapply) to construct a SNP × sample dosage matrix. Panel completeness was quantified as (i) per‐sample call rate, defined as one minus the proportion of missing genotypes across loci, and (ii) per‐locus call rate, defined as one minus the proportion of missing genotypes across samples. MAF was calculated from dosage values as $\mathrm{MAF}=\min (p,1-p),\mathrm{where}p=\frac{\overline{d}}{2}\mathrm{and}\overline{d}$ is the mean dosage across samples after excluding missing values Genotyping performance was visualized using histograms and empirical cumulative distribution functions of call rate, as well as MAF distributions for loci meeting standard usability thresholds (locus call rate ≥ 0.95; MAF ≥ 0.05), implemented using ggplot2.

2.8. Exploratory high coverage (50X)‐guided imputation to low coverage (2X)

This analysis represents an exploratory approach designed to evaluate how high‐coverage and low‐coverage datasets can be jointly utilized for downstream genomic studies. It was conducted independently of the enrichment panel and genetic diversity described above but demonstrates a framework for utilizing variant calls across sequencing depths. Variant representations between the high‐coverage panel (50X) and the low‐coverage cohort (2X) were harmonized against the red clover reference genome (T. pratense ARS_RC_1.1; RefSeq GCF_020283565.1). VCFs from both cohorts were left‐aligned and normalized using bcftools v1.18, and REF alleles were reconciled to the reference FASTA using the bcftools + fixref plugin. Sites with irreconcilable REF/ALT mismatches were discarded to prevent strand or allele orientation errors. After harmonization, the founder panel comprised 542,790 SNPs with 100% REF matches (no flips, swaps, or unresolved sites), while the 2X target dataset contained 8,205,365 SNPs prior to cross‐cohort alignment. To enable guided imputation and ensure allele consistency, the 2X dataset was restricted to sites identical to the founder panel (chromosome, position, REF, ALT). Genotype imputation was performed using Beagle v5.4 (build 22Jul22.46e), with the harmonized 50X founders as the reference haplotype panel and the intersected 2X dataset as the target.

Beagle generated genotype dosages and corresponding dosage R ² (DR²) accuracy estimates for each SNP. Post‐imputation filtering retained variants with DR² ≥ 0.8 and MAF ≥ 0.01 for downstream population analyses. Where permitted by statistical models, continuous dosage values were used instead of hard genotype calls. The final imputed dataset showed near‐complete concordance with the founder reference panel, negligible missingness, and consistent allele encodings across samples, supporting the robustness of this guided‐imputation strategy. Although this approach necessarily restricts analyses to sites shared between the founder and low‐coverage cohorts, thereby underrepresenting rare or private alleles unique to the 2X dataset, it maximizes allele consistency and enables reliable cross‐cohort comparisons. This trade‐off provides a scalable and cost‐effective framework for future T. incarnatum sequencing efforts at variable depths.

Imputation quality was assessed using Beagle's DR² metric. Site‐level DR² and AFs were extracted using bcftools +fill‐tags, and the proportion of variants exceeding key accuracy thresholds (DR² ≥ 0.8 and DR² ≥ 0.9) was quantified using standard Unix/AWK pipelines. Visualization was performed in Python using pandas v2.2 (McKinney, 2010) and matplotlib v3.8 (Hunter, 2007), generating DR² histograms, cumulative accuracy distributions, and DR²–AF scatter plots that comprehensively summarized imputation performance across the genome.

2.9. Assessment of reference genome mapping quality and panel coverage

To evaluate the evenness of enrichment panel coverage across the T. pratense reference genome, we assessed the relationship between mapping depth and probe density across genomic windows. Per‐sample mapping rates were extracted from SAMtools flagstat output files generated during the alignment step for all 45 genotypes in the 50X SNP‐discovery panel. Mean mapping depth per genomic window was computed for all 194 sequenced genotypes (45 at ∼50X and 149 at ∼2X) using mosdepth v0.3.10 (Pedersen & Quinlan, 2018) across 410 nonoverlapping 1‐Mb windows spanning the seven primary chromosomes (NC_060059.1–NC_060065.1). Windows were intersected with the final enrichment panel BED file using BEDTools intersect to count the number of probes per 1‐Mb window. The association between per‐window mean mapping depth and probe count was tested using Spearman rank correlation (two‐tailed, asymptotic p‐value approximation), with robustness confirmed by a permutation test (9999 permutations) and a trimmed analysis excluding windows flagged as outliers by the 3xIQR criterion.

2.10. Genetic diversity analyses

Ancestry composition among sampled genotypes was estimated using ADMIXTURE v1.3.0 by testing K values from 2 to 4, with the optimal K selected based on 10‐fold cross‐validation (CV) error (Alexander et al., 2009). Prior to analysis, LD pruning was performed using PLINK v1.9 with a 50‐kb sliding window, 5‐kb step size, and r ² ≤ 0.2, ensuring that highly correlated SNPs were removed before model‐based clustering. ADMIXTURE outputs were visualized in R using the ggplot2 and reshape2 packages (Wickham, 2007, 2016).

Principal component analysis (PCA) was conducted with the bigsnpr package (Privé et al., 2018), and hierarchical clustering was performed using the unweighted pair‐group method with arithmetic mean (UPGMA) based on Euclidean genetic distances computed with bigsnpr and bigstatsr. SNP density heatmaps and allele‐frequency distributions were generated using ggplot2, dplyr (Wickham et al., 2019), and ggrepel (Slowikowski, 2026) (Figure 2).

FIGURE 2.

Genome‐wide single‐nucleotide polymorphism (SNP) density heatmap of Trifolium incarnatum from the 50X high‐confidence SNP dataset, highlighting polymorphism hotspots across seven chromosomes. Darker orange regions represent higher SNP density, highlighting polymorphism‐rich genomic regions that may correspond to recombination or gene‐dense areas.

Genetic diversity analyses were based on a filtered SNP dataset derived from the 2X WGR panel of 149 T. incarnatum genotypes, which included only high‐confidence, biallelic sites with MAF ≥ 0.05 and missingness ≤ 0.10 following quality control and LD pruning. Genetic differentiation among regional groups was quantified by calculating pairwise fixation index [FST] values based on Nei's unbiased genetic distance estimator (Nei, 1987) using the genet.dist() function in hierfstat (Goudet, 2005). Global F‐statistics, including overall FST and inbreeding coefficient (FIS), were estimated using the wc() function in the same package following the method of Weir and Cockerham (1984). Observed and expected heterozygosity and inbreeding coefficients for each subpopulation were derived using basic.stats() in hierfstat. All plots were generated in R, primarily using ggplot2, with custom scripts adapted from Clark (2015) (https://github.com/lvclark/R_genetics_conv).

3. RESULTS

3.1. WGR and variant and SNPs discovery

WGR of 45 diverse crimson clover genotypes at ∼50X coverage produced a dataset with a mean mapping efficiency of 78.16% and per‐sample genome coverage ranging from 39.48X to 101.06X (mean = 65.44X). Reads were aligned to the T. pratense reference genome (423–437 Mb; Bickhart et al., 2022). From this high‐coverage dataset, ∼5.8 million total variants were identified, including approximately 4.96 million SNPs. After removing SNPs with >30% missing data, applying a MAF filter (≥0.05), and imputing missing genotypes with Beagle 5.4, 542,790 high‐confidence SNPs were retained (Table 1). Among the seven chromosomes, Chromosome 2 contained the greatest number of SNPs (93,614), while Chromosome 5 had the fewest (55,178) (Table 2). Mean SNP density was estimated at approximately 1347 SNPs per megabase, with density calculated in nonoverlapping 1 Mb windows across all chromosomes (Figure 2). Several regions exceeded 2500 SNPs per Mb, visualized as darker orange areas in Figure 2, highlighting genomic hotspots of elevated polymorphism. Notably, Chromosomes 1, 2, and 3 exhibited the highest SNP densities, suggesting the presence of marker‐rich regions useful for future genomic analyses.

TABLE 2.

Chromosomal distribution and single‐nucleotide polymorphism (SNP) density in crimson clover based on both high‐confidence and low confidence SNPs.

Chromosome	Size a (Mb)	SNP number
		∼50X coverage		∼2X coverage
		SNP number	SNP density per Mb	SNP number	SNP density per Mb
1	53.59	84,382	1574.6	353,540	6597.3
2	74.48	93,614	1256.9	402,557	5404.8
3	54.40	90,252	1659.1	389,206	7154.6
4	60.49	77,948	1288.5	370,493	6124.2
5	57.90	55,178	952.9	268,064	4629.2
6	46.12	64,766	1404.4	286,864	6220.6
7	59.21	76,650	1294.5	340,744	5754.7
Average	58.03	77,541.40	1347.27	344,495	5983.7

^a Based on the reference genome of red clover genome assembly ARS_RC_1.1 (Bickhart et al., 2022).

In parallel, a low‐coverage (2X) sequencing dataset comprising 149 additional genotypes were processed to evaluate the feasibility of cost‐effective genotyping for broader population analysis. Mapping efficiency averaged 82.19%, with genome coverage ranging from 0.67X to 6.98X (mean = 2.54X). Multiple variant‐calling pipelines identified 17,050,990 total variants, of which 16,051,990 were SNPs. After excluding SNPs with >30% missing data, the dataset was reduced to 11,282,126 SNPs; subsequent filtering with a MAF < 0.05 retained 3,913,665 SNPs. Following genotype imputation using Beagle 5.4 with haplotype phasing, 2,411,468 high‐confidence SNPs were retained in the final 2X dataset. Together, the high‐ and low‐coverage resequencing datasets provide a comprehensive genomic foundation for variant discovery, diversity assessment, and probe selection (Figure 1). This workflow integrates (a) JGI NGS data generation and storage, (b) Twist Bioscience hybrid‐capture synthesis for enrichment panel development, and (c) downstream application in breeding program(s).

3.2. Developing a SNP enrichment panel

We developed a custom hybrid‐capture SNPs panel containing initially approximately 30,000 loci. The panel integrated variants from three complementary sources: the JGI 50X high‐coverage dataset, the JGI 2X low‐coverage diversity dataset, and 6,803 validated SNPs from an independent Khufu low‐pass genotyping run conducted on Auburn crimson clover breeding germplasm in 2023 (Korani et al., 2021). All variant coordinates were standardized to the T. pratense ARS_RC_1.1 reference genome (RefSeq GCF_020283565.1) to ensure consistent chromosome indexing (NC_060059.1–NC_060065.1) and annotation across datasets. The combined dataset comprised 17.8 million raw SNPs that were subjected to a multistage refinement funnel to generate a manufacturable probe design. Site‐level quality control retained only biallelic variants with QUAL ≥ 30, MAF ≥ 0.03, and ≤10% missing data, producing approximately 365,000 high‐confidence SNPs. Strand‐ambiguous polymorphisms (A/T and C/G) were excluded to prevent probe‐orientation artifacts, reducing the dataset to ∼300,000 designable SNPs. To minimize redundancy, LD pruning was performed using a 50‐kb window, 5‐kb step, and r ² ≤ 0.2 (validated at r ² ≤ 0.1) to yield ∼119,000 largely independent loci. Variants were intersected with a merged genic annotation (“genic union”) encompassing gene bodies, coding sequences, and UTRs; overlapping records were resolved deterministically by prioritizing genic sites and higher AF.

Each retained variant was expanded ±60 bp (total 121 bp) to define capture intervals, and overlapping regions were merged to avoid redundant tiling. This process produced ∼66,000 nonredundant windows, approximately 84% of which intersected genic regions. Chromosome‐proportional representation was achieved by an even down‐selection of 24,000 intervals using a bin‐center heuristic, which partitioned each chromosome into equal‐length bins and selected the SNP nearest each bin midpoint. The resulting JGI‐derived design was unified with the previously analyzed data from our panel of crimson clover breeding germplasm and overlapping or adjacent targets were collapsed to yield 29,191 unique windows. A final top‐up step added 809 genic intervals, resulting in a final design containing approximately 30,000 capture probes. For transparency and reproducibility, gate counts were recorded at each stage of the design pipeline: 17,826,724 (merged SNPs), 364,742 (QC‐filtered), 300,246 (strand‐unambiguous), 119,129 (LD‐pruned), 66,149 (expanded ± 60 bp windows), 24,000 (even down‐select), 29,191 (JGI + Khufu SNP union), and 30,000 (final top‐up).This stepwise reduction illustrates the progression from raw whole‐genome variation to a balanced, design‐ready, mid‐density genotyping panel suitable for breeding applications. The finalized panel achieved balanced coverage across all seven primary chromosomes: Chr1 (NC_060059.1) = 3168, Chr2 (NC_060060.1) = 4400, Chr3 (NC_060061.1) = 3213, Chr4 (NC_060062.1) = 3574, Chr5 (NC_060063.1) = 3421, Chr6 (NC_060064.1) = 2725, and Chr7 (NC_060065.1) = 3499. Genic enrichment exceeded 75% for most chromosomes, peaking at 79.1% (NC_060061.1), with the lowest at 66.0% (NC_060063.1). Mean inter‐SNP spacing was ∼16.9 kb, with minimum spacing constrained to the designed ±60 bp threshold (122–123 bp). Density analyses in 1‐Mb bins showed ∼58–59 windows per Mb, with coefficients of variation ranging from 0.121 to 0.227, confirming uniform distribution across the genome (Figure 11).

FIGURE 11.

Genome‐wide distribution of the final 28,867 single‐nucleotide polymorphism (SNP)‐centered (±60 bp) capture targets in crimson clover.

Following design completion, the finalized capture panel was submitted to Twist Bioscience for independent in silico and manufacturing verification (Order Q‐547734; Design ID TE‐92745379). Twist's internal quality control confirmed successful synthesis and uniform probe performance across the Auburn Crimson Clover 30K hybrid‐capture panel, comprising 28,867 unique probes. All probes passed synthesis metrics, with >99.9% probe recovery, >98% achieving expected mapping efficiency, and normalized information entropy >0.92, indicative of balanced sequence complexity and hybridization efficiency. Mean sequencing coverage exceeded 30X, and GC distribution fell within the optimal manufacturing range for Twist's hybrid‐capture chemistry. These results collectively validate the design's uniformity, specificity, and manufacturability, confirming that the Auburn Crimson Clover 30K panel provides a reliable, genotyping platform for large‐scale precision genomics for crimson clover breeding.

3.3. Genotyping completeness and AF characteristics

To evaluate the genotyping performance of the Twist hybrid‐capture enrichment panel, we applied our Twist panel to 1039 breeding samples. Subsequent to sequencing, Khufu was used for variant calling and imputation as described in Section 2. The resulting imputed dataset (Figure 13) comprised of 72,181 SNPs. Genotyping completeness was uniformly high, with all samples and nearly all loci exceeding a 95% call‐rate threshold and most approaching complete genotyping (>99%), indicating minimal missing data following reference‐guided imputation.

FIGURE 13.

Sample‐ and single‐nucleotide polymorphism (SNP)‐level genotyping completeness and allele frequency characteristics of the imputed Twist enrichment dataset. (a) Empirical cumulative distribution function (ECDF) of sample call rates from the final imputed genotype matrix, with the dashed line indicating a 95% completeness threshold. (b) Jittered distribution of sample call rates across breeding lines. (c) Histogram of SNP‐level call rates showing locus‐level completeness across the enrichment panel. (d) Minor allele frequency (MAF) distribution of usable SNPs (call rate ≥ 95%), illustrating the frequency spectrum of retained markers. All metrics were calculated from the enriched‐only HapMap generated using Khufu low‐pass (∼3X) sequencing and reference‐guided imputation.

MAF analysis showed that 69.3% of loci were polymorphic at MAF ≥ 0.05, yielding 50,021 informative SNPs. The allele‐frequency spectrum was enriched toward intermediate frequencies (MAF 0.3–0.5), with mean and median MAF values of 0.239 and 0.236, respectively, indicating a marker set dominated by highly informative variant.

3.4. Reference genome mapping rates and enrichment panel coverage

Mapping rates across all 45 genotypes in the 50X SNP‐discovery panel averaged 79.51 ± 3.73% (range: 62.49%–93.16%) (Table S1). These rates are consistent with expectations for cross‐species alignment of an outcrossing diploid legume to a congeneric reference with a large, repeat‐rich genome (∼420 Mb, ∼50% repetitive content).

To assess the evenness of enrichment panel probe placement across the reference genome, mean mapping depth and probe count were compared across 410 nonoverlapping 1‐Mb genomic windows on the seven primary chromosomes (Table S2). A weak positive correlation was observed (Spearman ρ = 0.333, p < 0.001, n = 410 windows; permutation p < 0.0001, 9999 permutations), indicating that regions with higher mapping depth tended to harbor slightly more probes (Table S1). This relationship is expected, as regions of greater sequence conservation between T. incarnatum and T. pratense yield more discoverable variants during the SNP calling step. The correlation explains only 11.1% of the variance in probe density (ρ ² = 0.111), indicating that 89% of probe placement was determined by factors other than local mapping quality‐principally the bin‐center spacing algorithm and genic enrichment criteria. The chromosome‐proportional bin‐center heuristic used during panel design effectively constrained this effect: per‐chromosome mean probe density ranged from 45.2 probes per Mb (Chr5) to 56.3 probes per Mb (Chr6), and no 1‐Mb window across any chromosome lacked probes entirely (minimum = 10 probes per window; median = 53; IQR = 46–57). A trimmed analysis excluding 18 outlier windows (3XIQR criterion) yielded a similar correlation (ρ = 0.374, n = 392, p < 0.001), confirming that results were not driven by extreme values. These results demonstrate that while reference‐mapping quality had a modest influence on local SNP discovery, the panel design pipeline compensated by enforcing uniform genomic coverage across all chromosomes.

3.5. Geographic distribution and genetic diversity

Our diversity panel comprises 45 accessions sourced from North America, Europe, Asia, and Africa as depicted in Figure 3, capturing both cultivated and wild germplasm. Of these, 28 cultivated accessions including the cultivars Dixie, AU Sunrise, and UF Early White were obtained from the U.S. NPGS and UF. The remaining 19 accessions derive from wild‐collected populations across Europe. UPGMA dendrogram based on Euclidean genetic distances (Figure 4) showed moderate geographic grouping among the sampled genotypes. While some US and European genotypes grouped near one another, clear separation between regions was not observed, likely reflecting the species’ outcrossing nature and historical germplasm exchange. A few wild European genotypes, PI655002 (Hatif à Fleur Rouge), PI418900 (279), and PI591666 (93‐64), formed distinct branches apart from the main cluster, suggesting localized genetic divergence within European germplasm.

FIGURE 3.

Geographic distribution and origin of 45 crimson clover accessions spanning United States, Europe, Asia, and Africa.

FIGURE 4.

Unweighted pair‐group method with arithmetic mean (UPGMA) dendrogram showing genetic relationships among 45 crimson clover genotypes based on Euclidean distances derived from 542,790 high‐confidence single‐nucleotide polymorphisms (SNPs) (50X dataset). Each tip represents a single sequenced individual from a distinct accession, colored by geographic origin. Clustering reflects genetic similarity among sampled individuals and does not represent population‐level phylogenetic relationships.

3.6. Genetic diversity patterns among sampled genotypes

PCA was performed separately for the 50X and 2X datasets to explore genomic variation patterns in crimson clover. In the 50X dataset (Figure 5), PC1 explained 15.72% of the total genetic variance, while PC4 accounted for an additional 9.4%. Three 2D projections (PC1 vs. PC2, PC2 vs. PC3, and PC3 vs. PC4) revealed clustering among genotypes based on geographic origin (Figure 5) and improvement status (Figure 6). For the broader 2X dataset of 149 genotypes, PCA results are presented in Figures 7 and 8. PC1 (20.27%) and PC2 (12.93%) revealed clear separation by country, with US and European genotypes forming distinct clusters.

FIGURE 5.

Principal component analysis of 45 crimson clover genotypes from the 50X high‐confidence single‐nucleotide polymorphism (SNP) dataset, colored by geographic origin. Three 2D projections (PC1 vs. PC2, PC2 vs. PC3, and PC3 vs. PC4) capture the major axes of genetic variation among sampled individuals.

FIGURE 6.

Principal component analysis of 45 crimson clover genotypes from the 50X high‐confidence single‐nucleotide polymorphism (SNP) dataset, colored by improvement status, showing closer clustering of cultivars and breeding lines relative to wild and uncertain‐status materials.

FIGURE 7.

Principal component analysis of 149 crimson clover genotypes based on the 2X whole‐genome re‐sequencing (WGR) dataset, colored by geographic region. The scree plot (left) shows the proportion of variance explained by each of the first 10 principal components. Three 2D projections (PC1 vs. PC2, PC2 vs. PC3, and PC3 vs. PC4) illustrate the major axes of genetic variation among the sampled genotypes.

FIGURE 8.

Principal component analysis of 149 crimson clover genotypes from the 2X whole‐genome re‐sequencing (WGR) dataset, colored by improvement status. Format follows Figure 7, showing closer clustering of cultivars and breeding lines relative to wild and uncertain‐status materials.

Cultivars such as Dixie and certain AU and UF lines from the Southeastern United States clustered separately from European materials, indicating relatively cohesive genetic lineages. To estimate ancestry composition among the sampled genotypes, we ran ADMIXTURE with K values from 2 to 6 on the 2X dataset (∼2.4 million SNPs), using three replicate seeds per K. For each run, ADMIXTURE calculated 10‐fold CV error, which estimates model predictive error the lower the CV error, the better the model fits the data (Alexander et al., 2009). The lowest mean CV error was observed at K = 2 (0.52792 ± 0.00003), with only a slight increase at K = 3 (0.53386 ± 0.00005), suggesting that both K = 2 and K = 3 provide plausible descriptions of ancestry composition among the sampled genotypes (Figure 9a).

FIGURE 9.

Ancestry proportions of 149 crimson clover genotypes inferred with ADMIXTURE (K = 3) using the 2X whole‐genome re‐sequencing (WGR) dataset. Both panels display the same individuals and the same ancestry estimates using a single consistent color palette, sorted by (a) geographic region and (b) improvement status. These patterns illustrate the diversity of ancestry composition among sampled genotypes and do not represent population‐level structure.

At K = 2, genotypes grouped into two broad ancestry components, with most US cultivars and breeding materials dominated by a single cluster, while some European and Asian genotypes exhibited minor proportions of the second component. At K = 3, a third cluster emerged a low frequency, primarily among a few Asian and African genotypes, suggesting subtle geographic differentiation rather than discrete population structure. When sorted by improvement status, cultivars and wild materials were largely homogeneous for one dominant ancestry.

Complementary to the ADMIXTURE results, genetic differentiation was evaluated using the 2X coverage dataset consisting of 149 genotypes and approximately 2.4 million SNPs. Analyses of FST and FIS, computed following Weir and Cockerham's (1984) method, indicated that most genetic variation occurred within, rather than among, the sampled groups. Across geographic regions, differentiation was generally low, with FST = 0.0066 between the United States and Europe, 0.0117 between the United States and Africa, and 0.0172 between Europe and Africa. In contrast, Asian genotypes exhibited moderate divergence, showing FST = 0.062 from the United States, 0.0706 from Europe, and the highest differentiation from Africa (FST = 0.1537), suggesting partial geographic isolation of Asian lines.

Differentiation by improvement category was similarly minimal. Wild material and breeding lines were the most closely related (FST = 0.0019), while cultivars and breeding material showed the greatest separation (0.0088). All other comparisons, including wild versus cultivars (0.0027) and wild versus uncertain status material (0.0051), fell well below thresholds for moderate divergence, indicating that improvement activities have not substantially increased genetic stratification within crimson clover.

At the genome‐wide scale, the overall FST was 0.0105, confirming extremely low population differentiation among sampled genotypes. The inbreeding coefficient was negative (FIS = −0.0592), reflecting an excess of heterozygosity within groups. This pattern aligns with crimson clover's obligate outcrossing mating system and underscores that most genetic variation resides within rather than between the sampled groups.

In high‐confidence 50X coverage SNPs, Figure 10 illustrates the distribution of SNP alleles across four major geographic regions (United States, Asia, Africa, and Europe) alongside a bar chart (a) highlighting region‐specific SNP counts per chromosome. The Venn diagram reveals that the United States harbors the largest set of unique alleles (2778), followed by Asia (1061), Africa (1012), and Europe (197). A substantial core of 34,643 SNPs is shared among three of these regions, while another tri‐regional overlap comprises 60,524 SNPs, and 662 SNPs are exclusively shared between Asia and Africa, respectively. At the center, 37,892 SNPs are common to all four regions, underscoring a significant global genetic backbone. Meanwhile, the bar chart indicates that the United States consistently possesses more exclusive SNPs per chromosome than Europe. In addition; A second Venn diagram (Figure 10c) partitions SNPs according to improvement categories cultivars, breeding materials, uncertain improvement status, and wild materials (Figure 10). The large, shared core of 460,066 SNPs underscores a substantial common genetic base in crimson clover. However, each category also displays unique alleles, including 3215 SNPs in breeding materials, 675 in cultivars, 58 in uncertain lines, and 92 in wild materials. Cultivars have the highest number of exclusive SNPs across all chromosomes, followed by wild material, whereas breeding materials show fewer unique variants.

FIGURE 10.

Unique and shared single‐nucleotide polymorphisms (SNPs) among geographic regions and improvement categories in crimson clover based on the 50X high‐confidence SNP dataset, revealing both overlapping and exclusive allelic patterns.

4. DISCUSSION

4.1. WGR, genomics resources, and potential foundation for marker development

This study represents the first genome‐wide genomic assessment of crimson clover to date. Using high‐coverage (∼50X) WGR of 45 genotypes, we generated high‐resolution variant data spanning both coding and noncoding regions of the genome (Michael & VanBuren, 2020; Varshney et al., 2014). From this core dataset, we identified approximately 5.8 million variants, including ∼4.96 million SNPs. After stringent filtering, 542,790 high‐confidence SNPs were retained, providing a dense, robust marker resource.

These results compare favorably with previous GBS‐ and SSR‐based studies in related Trifolium species. For instance, red clover GBS analyses typically identified ∼12,500 high‐quality SNPs (Jones et al., 2020), while SSR surveys in white clover reported high allelic richness (PIC = 0.95) but lacked the genome‐wide resolution achieved here (Wu et al., 2021). By extending our strategy to include 149 additional genotypes sequenced at ∼2X genome‐wide coverage, we identified approximately 2.4 million high‐confidence SNPs with a mean PIC of 0.202. The 2X dataset effectively captured genetic diversity and ancestry patterns consistent with prior findings in Trifolium (Dias et al., 2008; Jones et al., 2020), while offering significantly greater genome coverage and variant density than GBS‐based approaches.

The 2X dataset yielded a greater number of variants, with approximately 17.05 million variants compared to 5.84 million identified in the 50X dataset. This difference is primarily due to the effects of joint variant calling across a much larger cohort (149 vs. 45 accessions). In joint genotyping, low‐coverage reads from multiple samples are evaluated simultaneously, allowing weak allelic signals to accumulate across individuals and increasing sensitivity to detect low‐frequency variants. As demonstrated by DePristo et al. (2011), this approach enables the recovery of the majority of common and rare variant sites even under shallow sequencing depth, as the collective evidence across samples compensates for per‐sample sparsity. This outcome aligns with the JGI resequencing strategy, in which each accession is represented by one deeply sequenced (50X) genotype to establish a high‐confidence reference for variant discovery and annotation, complemented by a broader low‐pass (2X) cohort designed to capture within‐species genetic diversity and family structure.

4.2. Deployment of WGR resources in crimson clover breeding

The genomic framework developed in this study provides a transferable foundation for accelerating improvement in open‐pollinated, obligate‐outcrossing species such as crimson clover. In parallel with genomic resource development, recurrent selection pipelines were initiated in 2022 to leverage the species’ high heterozygosity and capacity for natural recombination. Following the established model for open‐pollinated crop breeding (Viana et al., 2018; Yabe et al., 2014), the program begins with spaced plant trials (SPT), where individual plants are phenotyped for vigor, persistence, and reproductive fitness. The superior 5%–20% of plants are retained as female parents to form half‐sib families, which advance through progeny testing in a preliminary yield trial (PYT). Top‐performing maternal lines are combined into balanced bulks and subsequently evaluated in ALT across environments and years, allowing estimation of genetic variance and genotype‐by‐environment interactions. Recurrent selection improves the population by identifying superior individuals and recombining them each cycle, allowing gradual increases in mean performance while maintaining the genetic variation needed for long‐term selection (Annicchiarico, 2003; Brummer, 1999). Our future breeding strategy will leverage genomic selection at the SPT stage using yield data from progeny testing trials (PYT, advanced yield trial [AYT]). We will integrate phenotypic data from SPT, PYT, and AYT data with application of our targeted enrichment panel to enable genomic prediction and to maximize the value of each data point.

We have developed a custom (Figure 11) 28,867 Twist hybrid‐capture panel specifically for crimson clover. The design incorporated 28,931 probes spanning 24,106 merged genomic regions, with 21,051 regions successfully tiled and an overall probe coverage of 84.3% (3.14 Mb of sequence targeted). Importantly, over 75% of probes target genic regions, with inter‐probe spacing averaging ∼16–17 kb, ensuring both functional relevance and uniform genome‐wide coverage. This platform is already being applied to genotype Auburn's recurrent selection families, providing marker data for both forward selection of superior individuals and retrospective analysis of AF changes across cycles. Such a mid‐density platform bridges genome‐wide variant discovery with operational breeding, offering immediate utility for our program. The wealth of SNPs discovered through WGR of crimson clover provides the discovery layer that enabled the design of our enrichment panel. While reduced‐representation approaches such as GBS, RAD‐seq, and targeted SNP capture remain attractive for their scalability and cost‐effectiveness (Elshire et al., 2011), their resolution has historically been limited. For example, prior GBS studies in Trifolium recovered only ∼12,500 high‐confidence SNPs in red clover after stringent filtering (Jones et al., 2020), highlighting the relative power of WGR‐based discovery. By contrast, our dataset retained more than 542,790 high‐confidence SNPs, a resource large enough to support robust, genome‐wide marker development. Leveraging this foundation, we developed a 30K Twist hybrid‐capture panel for crimson clover, placing this species on par with other legumes where mid‐ to high‐density genotyping resources have transformed breeding. In soybean, the SoySNP50K BeadChip underpins diversity assessment, GWAS, and genomic prediction (Lee et al., 2015; Song et al., 2013); in alfalfa, capture‐based and array platforms ranging from ∼9K SNPs have dissected biomass and quality traits from identified 900K SNP (Annicchiarico et al., 2015; X. Li et al., 2014) and in peanut (Arachis hypogaea L.), the Axiom_Arachis2 SNP array enabled high‐resolution trait mapping across breeding programs (Clevenger et al., 2017). Our enrichment panel thus delivers the same scale and reproducibility, while being tailored to an obligately outcrossing forage legume with limited prior genomic resources. Twist Bioscience provides synthesis and enrichment services for the crimson clover 30K hybrid‐capture panel; inquiries regarding availability should be directed to Paul Doran (pdoran@twistbioscience.com).

Our crimson clover panel shares the same design logic as these efforts (Figure 12): starting from WGR‐derived discovery sets, then filtering for quality, functional relevance, and genome coverage. With 30K evenly spaced and largely genic probes, this resource provides breeders with resolution comparable to soybean and peanut arrays, while being tailored to a primarily outcrossing forage legume with limited prior resources. The panel comprised 28,867 validated probes, all passing Twist Bioscience's synthesis and hybridization quality metrics. With >99.9% probe recovery, >98% mapping efficiency the panel, demonstrates high manufacturing fidelity and capture uniformity suitable for reproducible, large‐scale genotyping. Although probe specificity was verified in silico against the T. pratense ARS_RC_1.1 reference genome, it is important to note that structural or segmental duplications unique to the T. incarnatum genome could potentially introduce minor off‐target hybridization not detectable in the reference‐based evaluation. Such effects are expected to be limited in scope, given the overall probe spacing and GC‐balance, but future mapping to a de novo T. incarnatum assembly will allow empirical assessment and refinement of panel specificity. Moreover, this effort positions crimson clover among the growing number of legumes supported by standardized, mid‐density genotyping platforms for predictive breeding applications such as genomic selection and genomic prediction (Concepcion et al., 2025), MAS, and long‐term diversity monitoring. As the ongoing assembly of a high‐quality T. incarnatum reference genome becomes available, these reads and SNP coordinates will be realigned to refine probe specificity, variant annotation, and cross‐species comparability with T. pratense, further improving the accuracy of future mapping and genomic prediction studies.

FIGURE 12.

Design and validation pipeline of the crimson clover 28,867‐single‐nucleotide polymorphism (SNP) enrichment panel for public use.

Beyond SNP arrays, our genomic resource can also be adapted to flexible, lower cost systems like DartTag (diversity arrays technology) and kompetitive allele‐specific PCR assays offered by LGC Genomics. Both approaches rely on high‐quality, validated SNP catalogs for assay design. The 542K high‐confidence SNP set and the ∼28K enrichment panel provide an ideal substrate for selecting subsets of markers for trait‐specific genotyping. Integrating high‐quality SNPs into genotyping arrays or amplicon panels is expected to greatly improve trait mapping resolution and selection accuracy. It is important to emphasize, however, that directly combining 50X and 2X coverage datasets for genetic diversity analyses may introduce coverage bias and skew AF estimates. This was evident in our preliminary PCA analyses that showed separation of 50X and 2X samples, likely due to read depth discrepancies rather than true genetic differentiation. As such, diversity metrics, ancestry estimation (e.g., ADMIXTURE), and FST should rely on low‐pass (2X) samples exclusively to avoid coverage‐driven artifacts.

4.3. Utility of a mid‐density enrichment panel for predictive breeding

We evaluated a custom Twist hybrid‐capture panel in our crimson clover panel of breeding materials using Khufu low‐pass sequencing with reference‐guided imputation. The resulting imputed (I)HapMap achieved consistently high sample‐ and locus‐level call rates (Figure 13), yielding a near‐complete genotype matrix appropriate for breeding‐scale analyses. Importantly, these metrics reflect the final imputed genotypes used in downstream inference, rather than raw read depth or capture‐on‐target rates. The allele‐frequency spectrum is consistent with predictive breeding needs: enrichment of intermediate‐frequency variants (MAF = 0.3–0.5) increases information content for ancestry estimation and genomic prediction, while retention of lower‐frequency alleles preserves founder‐haplotype diversity and supports monitoring allele‐frequency trajectories across selection cycles. Relative to fixed SNP arrays, sequence‐based hybrid capture is modular and updatable, whereas arrays are constrained by static marker content and can exhibit ascertainment bias driven by discovery panels and design filters that distort allele‐frequency spectra and downstream LD/FST‐based inference (Albrechtsen et al., 2010; Lachance & Tishkoff, 2013). Benchmarking in animal breeding further supports capture as an array‐comparable platform: a Twist capture panel in cattle produced ∼98% genotype concordance, R ² = 0.98, relative to SNP arrays, confirming suitability for genomic selection pipelines (Ren et al., 2024). Compared with restriction‐enzyme‐based GBS, hybrid capture concentrates sequencing effort on predefined loci, improving locus stability and reducing dropout, an advantage for routine, cycle‐to‐cycle breeding operations that require consistent marker sets (Poland et al., 2012). The availability of 50,021 informative common markers (69.3% of loci polymorphic at MAF ≥ 0.05) provides a practical foundation for breeding applications. This position the crimson clover Twist panel as a practical mid‐density genotyping platform bridging WGR‐based SNP discovery and operational predictive breeding in an obligately outcrossing forage legume.

4.4. Insights from low‐coverage (2X) imputation of high‐coverage (50X) data: A low‐cost and scalable pipeline for future generations

Recognizing the obligately outcrossing and highly heterozygous nature of crimson clover, we outline a two‐tier sequencing framework designed to balance cost efficiency, genomic resolution, and cross‐generational continuity within an active breeding program. This framework combines high‐coverage whole‐genome sequencing (50X) of selected founder accessions from the 194 individuals that established the Auburn crimson clover breeding program in 2022–2023 (C₀) with planned low‐pass resequencing (∼2X) of subsequent breeding cycles (e.g., C₁–C₃). This design follows the low‐coverage imputation paradigm of Pasaniuc et al. (2012), which demonstrated that accurate genotype recovery can be achieved from sparse sequencing when imputed against a dense haplotype reference panel, providing a scalable and economical alternative to conventional high‐depth resequencing. The 50X founder sequences generated a high‐confidence SNP discovery panel containing approximately 542,790 variants, which serves as a haplotype reference scaffold for future imputation of low‐pass data.

After stringent cross‐cohort harmonization (coordinate normalization, REF/ALT alignment, and exclusion of ambiguous polymorphisms), 488,078 loci were retained as candidates for imputation. Using Beagle v5.4, we evaluated imputation performance by imputing low‐pass genotypes against the founder haplotypes, achieving a mean imputation accuracy of DR² = 0.95 (Figure 14), with downstream analyses restricted to DR² ≥ 0.8 and MAF ≥ 0.01. This validation demonstrates the feasibility of applying low‐pass sequencing with imputation as a routine genotyping strategy in future breeding cycles, enabling genome‐wide marker coverage for GWAS, genomic prediction, and selection decisions at substantially reduced cost. Because crimson clover currently lacks a species‐specific reference genome, potential mapping or representation biases were minimized through conservative site filtering and cross‐species harmonization to the T. pratense assembly. Importantly, this framework is extensible to sub‐1X coverage (∼0.8X), allowing future breeding populations to be genotyped efficiently while maintaining high accuracy for common variants (MAF ≥ 0.05). Embedding this approach within each selection cycle will enable continuous, genome‐wide genotyping to inform long‐term breeding population improvement and cultivar development.

FIGURE 14.

Imputation accuracy and allele‐frequency relationship for the 2X crimson clover dataset imputed to the 50X founder reference panel. (a) Scatterplot of Beagle dosage R ² (DR²) versus minor allele frequency (MAF) showing consistently high imputation accuracy across the allele‐frequency spectrum, with only modest reduction among rare variants (MAF < 0.05). (b) Histogram of DR² values for all 542,790 imputed single‐nucleotide polymorphisms (SNPs) illustrating a strong right‐skewed distribution and a predominant peak at DR² = 1.0, indicating near‐perfect genotype recovery across most sites. (c) Boxplot of DR² (0–1.0) values across MAF bins (full range), showing substantially reduced accuracy for ultrarare variants (MAF < 0.01) but uniformly high accuracy (median DR² = 1.0) for all variants with MAF ≥ 0.01. (d) Zoomed MAF‐binned (DR² (0.7–1.0) distributions for MAF ≥ 0.01, highlighting consistently high and stable imputation accuracy across common and intermediate‐frequency alleles. Together these plots demonstrate the reliability of high‐coverage‐guided imputation for crimson clover, supporting the use of low‐pass sequencing (2X) as a cost‐efficient yet accurate genotyping strategy for downstream genomic analyses and breeding applications.

4.5. Insights into genetic diversity among sampled genotypes

Genetic diversity analyses revealed that crimson clover germplasm from the United States and Europe share a common genetic background. Similar observations were made in red clover and white clover, where most genetic variation resides within rather than between sampled groups (Dias et al., 2008; Wu et al., 2021). The relatively tight clustering of US commercial cultivars and broader spread of European wild types observed in our PCA and UPGMA results mirrors the modest divergence found in red clover (Jones et al., 2020). From the UPGMA dendrogram (Figure 4), several genotypes stood out, including three genetically distinct European wild types, PI655002 (Hatif à Fleur Rouge), PI418900 (279), and PI591666 (93‐64), as well as CHI‐1, a breeding material from Texas showed distinctness compared to all other genotypes. This highlights the capacity of crimson clover wild relatives to add diversity to breeding programs, as breeding programs based on a narrow genetic base may lack variation for traits such as drought tolerance that will be important for future climate adaptation. This pattern echoes findings in white clover, where geographic separation and local adaptation produced by genetically distinct groups (Wu et al., 2021). These outliers may warrant further phenotypic investigation. PCA also (Figure 5) confirmed tight clustering among AU cultivars such as Auburn and AU Reseeding. Notably, both sources of AU Robin, one from NPGS and one commercial, grouped closely. Overall, these findings are consistent with previous studies in related Trifolium species, where population structure was similarly resolved using lower density markers. For instance, geographic clustering was observed in white clover using SSR markers and in red clover using GBS data (Jones et al., 2020; Wu et al., 2021).

The identification of private SNPs within our high coverage 50X dataset, 675 in cultivars, 92 in wild materials, and 13 in breeding materials, underscores the capacity of WGR to detect lineage‐specific alleles. Moreover, our finding of 2778 SNPs unique to US genotypes and 197 to European lines, while these SNPs are not necessarily functional, their distribution provides insight into patterns of genetic uniqueness across improvement categories.

ADMIXTURE analysis at K = 3 revealed weak but biologically consistent ancestry differentiation among the sampled genotypes panel (Figure 9). At this level, most European and US genotypes were dominated by a single major ancestry component, while the secondary and tertiary components appeared only at low frequency across the panel. Although one minor component was more visible in a subset of Asian and African accessions, these contributions were modest, and neither region formed a distinct cluster; individuals from Europe and the United States also occasionally showed membership in the same minor components. This pattern indicates limited continental differentiation and substantial shared ancestry among regions. When grouped by improvement status, cultivars and breeding materials were nearly homogeneous for a single cluster, whereas wild and unclassified genotypes displayed more visibly mixed ancestry profiles, reflecting retention of broader ancestral variation. Pairwise FST (<0.02 among most regions; overall = 0.0105) and negative FIS (−0.0592) corroborate extensive gene flow and excess heterozygosity typical of obligate outcrossers. Together, these results support crimson clover as a broadly panmictic species with weak geographic differentiation among sampled genotypes but detectable signals of domestication‐related genetic consolidation.

Pairwise FST values confirm low‐to‐moderate genetic differentiation across regions. Our findings, FST values <0.02 for most interregional comparisons, and <0.07 between Asia and other regions, align with those reported for red clover (Jones et al., 2020) and white clover (Wu et al., 2021), supporting the concept of T. incarnatum as a panmictic species with weak differentiation. The overall genomic diversity FST = 0.0105 and FIS = −0.0592 calculated via Weir and Cockerham's method further validate this interpretation and parallel the excess heterozygosity seen in obligate outcrossers like T. repens (Wu et al., 2021).

Our results also parallel the only prior molecular diversity study in crimson clover by Steiner et al. (1998), who analyzed 40 accessions using 161 AFLP markers and reported a narrow genetic base among US cultivars. Similarly, our PCA and ADMIXTURE analyses confirm strong clustering among US cultivars and a broader spread among European and wild accessions, consistent with a shared genetic origin and subsequent breeding bottlenecks. However, by leveraging WGR, we extend the resolution and scope of that foundational work from hundreds of anonymous AFLP fragments to 2.95 million high‐confidence SNPs across a globally representative panel of 194 genotypes. This dataset captures both within‐accession and among‐accession variation, enabling finer‐scale detection of private SNPs, subtle ancestry variation among wild and uncertain‐status materials, and regional differentiation that was previously unresolved. Collectively, these results validate the broad patterns reported by Steiner et al. while providing a comprehensive, genome‐wide framework for modern genetic and pre‐breeding applications in T. incarnatum.

4.6. Reference genome limitations and future directions

A limitation of the current study is the alignment of all WGR data to the red clover reference genome (ARS_RC_1.1), necessitated by the absence of a T. incarnatum genome assembly. Mapping rates across the 45 genotypes in the 50X SNP‐discovery panel averaged 79.51 ± 3.73%, indicating that approximately 20% of sequencing reads could not be aligned, likely reflecting repeat‐mediated mapping failures, structural differences between the two species, and lineage‐specific sequences absent from the reference. To assess whether this unmapped fraction affected the enrichment panel design, we examined the relationship between mean mapping depth and probe density across 410 nonoverlapping 1‐Mb genomic windows on all seven primary chromosomes. A weak positive correlation was observed (Spearman ρ = 0.333, p < 0.001, n = 410), indicating that regions of greater interspecific conservation yielded slightly more discoverable variants and consequently more probes. However, this effect was small (ρ ² = 0.111) and was effectively constrained by the chromosome‐proportional bin‐center heuristic: per‐chromosome mean probe density ranged from 45.2 to 56.3 probes per Mb, and no 1‐Mb window lacked probes entirely (minimum = 10; median = 53). These results indicate that probe coverage is uniform across the mappable portion of the reference genome, though genomic regions unique to T. incarnatum that lack homology to T. pratense remain unrepresented in both the variant catalog and the enrichment panel. To address this limitation directly, we have initiated a haplotype‐resolved, chromosome‐scale genome assembly for crimson clover. This species‐specific reference will enable realignment of all WGR reads and panel probe coordinates, more accurate variant discovery, structural variant detection, comprehensive annotation of genes and regulatory elements relevant to agronomic traits, and empirical quantification of any genomic regions currently underrepresented due to cross‐species reference mapping.

5. CONCLUSION

This study delivers the first genome‐wide genomic resource for crimson clover and establishes a practical foundation for both modern breeding and genebank‐enabled crop improvement. Through WGR of a globally representative panel, we generated a high‐confidence catalogue of genome‐wide variation that far exceeds the resolution of previous marker‐based studies. These data characterize genetic diversity among sampled genotypes, capture rare and private alleles, and provide the most complete view of standing genetic diversity within the species.

From a breeding perspective, these resources position crimson clover from a historically under‐resourced species to one that can fully participate in predictive and data‐driven improvement pipelines. The development of a publicly available mid‐density (28,867 SNPs) hybrid‐capture enrichment panel bridges genome‐wide variant discovery with operational breeding. The panel has been synthesized, validated, and deployed within the Auburn crimson clover breeding program, enabling direct integration of genome‐wide marker information into recurrent selection pipelines. In addition to supporting genomic prediction, this platform enables tracking of allele‐frequency changes across selection cycles and more effective management of genetic diversity, a key requirement for sustained genetic gain in open‐pollinated, highly heterozygous species.

Equally important is the value of these resources for gene banks and germplasm management. The genome‐wide SNP dataset enables robust characterization of accessions, identification of genetically distinct or underrepresented materials, and detection of redundancy within collections. The availability of standardized and transferable markers further supports interoperability between gene banks and breeding programs, facilitating more effective utilization of conserved diversity. These advances align crimson clover with the genotyping platforms and SNP panel resources already established in soybean, alfalfa, and peanut, bringing this species to parity with other well‐resourced legumes and enabling long‐term, genomics‐enabled improvement.

AUTHOR CONTRIBUTIONS

Mark Philip Castillo: Conceptualization; data curation; formal analysis; investigation; methodology; resources; validation; visualization; writing—original draft; writing—review and editing. Oluwaseye Gideon Oyebode: Conceptualization; data curation; writing—review and editing. Jayson Talag: Investigation; resources. Victoria Bunting: Resources. Navneet Kaur: Investigation; resources. Paul Doran: Resources; validation. Kerrie Barry: Funding acquisition; resources. Dario Competti: Investigation; resources. Jeremy Schmutz: Data curation; funding acquisition. Brandon Schlautman: Funding acquisition. Alan Humphries: Funding acquisition; writing—review and editing. Kioumars Ghamkhar: Funding acquisition; writing—review and editing. Virginia Moore: Funding acquisition; writing—review and editing. Esteban Rios: Funding acquisition; resources; writing—review and editing. Alex Harkess: Funding acquisition; resources; writing—review and editing. Marnin Wolfe: Conceptualization; funding acquisition; investigation; project administration; resources; supervision; writing—review and editing.

CONFLICT OF INTEREST STATEMENT

Paul Doran is an employee of Twist Bioscience, which manufactures and commercially distributes the crimson clover hybrid‐capture enrichment panel described in this study. The remaining authors declare no conflicts of interest.

Supporting information

Attached bed file format we included for enrichment panel (CrimsonClover30K_AuburnUniversity.bed), upon request is the high and low coverage and reference panel imputed SNPs (due to huge size), email mbc0078@auburn.edu or mpcgenomics@gmail.com.

Castillo, M. P. , Oyebode, O. G. , Talag, J. , Bunting, V. , Kaur, N. , Doran, P. , Barry, K. , Competti, D. , Schmutz, J. , Schlautman, B. , Humphries, A. , Ghamkhar, K. , Moore, V. , Rios, E. , Harkess, A. , & Wolfe, M. (2026). Development and deployment of a public mid‐density genotyping panel for crimson clover (Trifolium incarnatum L.) based on whole‐genome re‐sequencing of diverse germplasm. The Plant Genome, 19, e70297. 10.1002/tpg2.70297

DATA AVAILABILITY STATEMENT

Raw whole‐genome sequencing data generated in this study have been deposited in the NCBI Sequence Read Archive (SRA) under the umbrella BioProject accession PRJNA1402467 (https://www.ncbi.nlm.nih.gov/bioproject/1402467), which links all associated data‐level BioProjects corresponding to individual accessions. Processed variant datasets and resources related to the mid‐density genotyping enrichment panel are provided in the Supporting Information. Additional processed datasets supporting the conclusions of this study are available from the corresponding author upon request.