Using genomic selection to examine subgenome dominance and epistasis in allopolyploid strawberry

Abstract

Allopolyploids are organisms that possess multiple sets of chromosomes derived from distinct ancestral species, resulting in multiple subgenomes. Many important crops are allopolyploid, including wheat (Triticum aestivum), cotton (Gossypium hirsutum), coffee (Coffea arabica), and strawberry (Fragaria × ananassa). In allopolyploids, subgenome dominance often emerges, with strong influence on gene expression, epigenetic regulation, and gene conservation. Subgenome dominance has been extensively characterized at the molecular level but very little at the phenotypic level. In this study, we investigated the importance of subgenome interactions for predicting nine phenotypes using a large dataset of 6718 genotypes from the University of Florida strawberry breeding population. We tested multi‐kernel genomic selection models accounting for subgenome and epistasis effects to test the genetic contribution of each ancestral strawberry subgenome. Across three yield‐related phenotypes, two fruit quality phenotypes, and four disease resistance phenotypes, the contributions of the four subgenomes were highly variable. On average, subgenome B contributed the most genetic variation, followed by subgenome A, then subgenome C, and lastly subgenome D. Using genomic selection models with epistasis kernels, we estimate that epistasis contributed between 12% and 46% of genetic variation for the nine phenotypes. Lastly, we used two different cross‐validation scenarios to show that genomic selection models incorporating subgenome and epistatic effects improve predictive ability by 1%–34% depending on the phenotype and the genomic selection training set.

Core Ideas

• Subgenome effects were estimated in octoploid strawberry, revealing that subgenome dominance is phenotype dependent.

• Genomic selection models with subgenome and epistasis kernels improved multi‐season predictions.

• Across nine strawberry phenotypes, epistasis accounted for 12%–46% of genetic variation.

Plain Language Summary

Allopolyploids are organisms that inherit multiple sets of chromosomes from more than one ancestral species. Thus, allopolyploid species possess multiple subgenomes that usually differ in how much they retain and express genes. In this study, we investigated the importance of four different subgenomes in cultivated strawberry for predicting nine traits in 6000 strawberry varieties and selections from the University of Florida. We used genomic selection models accounting for subgenome effects and their interactions to determine the level of contribution from each ancestral strawberry subgenome. We observed highly variable subgenome contributions to each trait. In general, subgenomes A and B contributed the most genetic variation, followed by C and then D. Additionally, we demonstrated that accounting for subgenome effects and genetic interactions improves the ability to predict the best‐performing parents and varieties in strawberry breeding.

Abbreviations

AUDPC

area under the disease progress curve

CCR

colletotrichum crown rot

CR

charcoal rot

DIC

deviance information criterion

PhCR

phytophthora crown rot

PM

powdery mildew

RCB

randomized complete block

SNP

single‐nucleotide polymorphism

1. INTRODUCTION

Polyploidy is an evolutionary force that is widespread among eukaryotes and has played a key role in speciation and diversification (Van de Peer et al., 2017). Generally, there are two evolutionary paths to polyploidy: autopolyploidy and allopolyploidy. Autopolyploids achieve polyploidy by doubling chromosomes within a species, whereas allopolyploids achieve polyploidy by doubling chromosomes through the hybridization of multiple species. After an allopolyploid is created, rapid genetic and epigenetic changes occur in order to stabilize the genome (Bird et al., 2018). Subgenome dominance is one of the mechanisms that has arisen to stabilize polyploid genomes (Alger & Edger, 2020), resulting in unequal rates of gene conservation, transcription, and transposable elements across the subgenomes. Subgenome dominance has now been reported in maize (Schnable et al., 2011), Arabidopsis (Thomas et al., 2006), wheat (Li et al., 2014), Brassica rapa (Cheng et al., 2016), bamboo (Ma et al., 2024), and strawberry (Edger et al., 2019; Fang et al., 2024; Han et al., 2025; Song et al., 2024).

Cultivated strawberry (Fragaria × ananassa) is a fruit crop grown worldwide for its high yield, pleasant taste, and nutritional properties. Strawberry is a hybrid species between two allo‐octoploid species (Fragaria virginiana and Fragaria chiloensis) (Darrow, 1966). Recent advancements in genomics have illustrated the complex nature of its genome (Fan & Whitaker, 2024; Feng et al., 2021; Hardigan, Feldmann, et al., 2021; Jin et al., 2023). A proposed chromosome nomenclature describes the strawberry genome in four subgenomes (A, B, C, and D) composed of seven chromosomes each (Hardigan, Lorant, et al., 2021). There is ample evidence that the A subgenome belongs to the diploid progenitor Fragaria vesca and that the B subgenome belongs to Fragaria iinumae (Hardigan, Lorant, et al., 2021; Jin et al., 2023). The progenitors of the C and D subgenomes have been debated, but the most likely candidates are extinct diploid species that are closely related to F. iinumae (Hardigan, Lorant, et al., 2021; Jin et al., 2023). Recent analysis using k‐mer clustering of transposable elements has provided a more concrete subgenome assignment for all chromosomes (Session & Rokhsar, 2023). Despite the open questions around the origins of the C and D subgenomes, the cultivated strawberry genome has long demonstrated disomic segregation and high karyotypic stability across a wide range of genetic diversity (Hardigan et al., 2020).

Several reports have indicated that subgenome A (F. vesca) is dominant over the other three subgenomes in strawberry (Edger et al., 2019; Fang et al., 2024; Han et al., 2025; Song et al., 2024). For example, Edger et al. (2019) reported less gene loss in subgenome A and a significant bias in homoeolog expression toward the A subgenome. The authors noted that this was especially true for homoeologs responsible for fruit traits such as anthocyanin biosynthesis and specific flavor compounds (Edger et al., 2019). More recently, higher levels of gene conservation in subgenome A was also reported in a reference quality genome of the University of Florida variety “Florida Brilliance” (Han et al., 2025). Fang et al. (2024) demonstrated a significantly higher number of accessible chromatin regions per gene in subgenome A using MNase‐hypersensitivity sequencing (MNase‐seq). This result was supported by higher levels of homoeolog expression in subgenome A and preferential sequence conservation of accessible chromatin regions (Fang et al., 2024). The authors used gene ontology to suggest that subgenome A was particularly dominant for stress‐related response genes (Fang et al., 2024). Another study supported the findings from Edger et al. (2019) and Fang et al. (2024), reporting higher levels of expression and retention rates of genes in subgenome A (Song et al., 2024). This study also demonstrated that subgenome A exhibited much lower methylation levels and higher expression of fruit maturity genes (Song et al., 2024). Despite the substantial evidence for subgenome dominance at the molecular level in strawberry, we are not yet able to link these findings with phenotypic variation.

The use of high‐quality, high‐throughput genotyping data have made large‐scale genotype‐to‐phenotype experiments feasible in plants. To make use of this new scale of genotypic data, genomic selection was proposed as a method for estimating the genetic merit of individuals using genomic information (Meuwissen et al., 2001). Most genomic selection models have been developed for diploid species, focusing on additive genetic effects. There is a need to evaluate and adapt these methods for polyploid species. Santantonio et al. (2019), for example, outlined the statistical framework needed to calculate within‐ and across‐subgenome effects in wheat. The authors reported different subgenome contribution patterns for yield, test weight, plant height, and heading date in allohexaploid wheat (Santantonio et al., 2019). In the same species, Cuevas et al. (2024) reported dominant levels of genetic variation for each subgenome (A, B, and D) corresponding to three different wheat disease traits. Additionally, the B wheat subgenome demonstrated significantly more genetic variation for milling traits (Bernardo, 2021). On the other hand, other authors reported similar levels of genetic contribution were assigned to each wheat subgenome for grain yield in separate wheat populations (Bernardo, 2021; Tessele et al., 2025). Altogether, these studies suggest that subgenome dominance in wheat is trait‐dependent and may not be consistent across different classes of phenotypes (i.e., disease resistance or milling quality). To our knowledge, the use of genomic prediction for subgenome modeling has thus far been done exclusively in wheat.

The primary interest of this study was to apply genomic selection models in strawberry that incorporate subgenome and epistasis kernels. Here we used a large strawberry breeding dataset spanning a decade to examine (1) the contributions of each strawberry subgenome to several phenotypes in a breeding program; (2) the complex genetic architecture of several strawberry phenotypes; and (3) the efficacy of genomic selection models with additive genomic effects, additive subgenome effects, epistasis effects, subgenome epistasis effects, and subgenome covariance effects.

2. MATERIALS AND METHODS

2.1. Breeding material

The University of Florida strawberry breeding program is a recurrent selection program utilizing a partial diallel mating scheme (Osorio et al., 2021). Each season, the breeding program makes new crosses, evaluates these families the following season using visual selection, and advanced individuals are then tested in yield trials the next season. The total dataset used in this study represents 6718 individuals genotyped and phenotyped across 10 consecutive growing seasons spanning 2013–2023. This dataset included two primary phenotyping experiment types: yield phenotyping trials and pathology trials. The yield phenotyping trials consisted of 4930 advanced strawberry breeding selections and cultivars derived from 660 unique cross combinations (biparental families) arising from 313 parents. The average number of full sibs in each family was eight, and the median was four. For a given genotype, the median number of full‐sibs was six and the median number of half‐sibs was 69 within a season, whereas the median number of full‐sibs was zero and the median number of half‐sibs was 82 across seasons. The pathology trials consisted of 6371 unique genotypes that were represented by a combination of advanced selections and unselected biparental families. Further description of the genotypes tested for each disease trait is recorded below and summarized in Table 1.

TABLE 1.

The number of non‐overlapping genotypes within each season for nine strawberry phenotypes.

Phenotype	Season										Total
	2013–2014	2014–2015	2015–2016	2016–2017	2017–2018	2018–2019	2019–2020	2020–2021	2021–2022	2022–2023
Average weight	799	867	318	383	393	404	423	437	451	447	4922
Brix	799	869	317	383	393	404	423	437	448	447	4920
Culls	799	869	321	386	393	404	423	437	451	447	4930
Firmness	578	582	307	382	393	404	423	437	448	447	4401
Marketable yield	799	869	319	386	393	404	423	437	451	447	4928
PhCR	773	898	284	313	370	396	393	392	0	0	3819
CR	0	0	0	0	542	477	180	238	130	371	1938
CCR	579	97	252	104	0	88	145	145	107	0	1517
PM	0	0	0	0	272	503	468	487	0	99	1829

Note: This number coincided with the size of the validation set used in genomic selection delete‐season cross‐validation. For each phenotype, the number of genotypes used for estimating phenotypic variance components, subgenome effects, and genetic architecture is presented as “Total”.

Abbreviations: CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

Core Ideas

• Subgenome effects were estimated in octoploid strawberry, revealing that subgenome dominance is phenotype dependent.

• Genomic selection models with subgenome and epistasis kernels improved multi‐season predictions.

• Across nine strawberry phenotypes, epistasis accounted for 12%–46% of genetic variation.

2.2. Phenotypic analyses for yield and fruit quality

The experimental design remained constant over the course of the 10 growing seasons represented in this study. In each season, all experiments were grown as a randomized complete block (RCB) design at the Gulf Coast Research and Education Center (GCREC) in Balm, FL (lat. 27°45′37.98″ N, long. 82°13′32.49″ W).

The yield trials were grown as single‐plant plots and were grown in either a five‐replication RCB or a three‐replication RCB. A five‐replication RCB was grown in all 10 seasons of phenotyping, and an additional three‐replication RCB was grown in 2013–2014, 2014–2015, and the 2015–2016 season. The three‐replication and five‐replication trials were grown side‐by‐side with identical agronomic conditions. Over the 10 seasons, some replications were grown as a single bed and some were grown as two beds. The plants were planted in two offset columns within each bed. A more robust explanation of the trialing beds is described by Zheng et al. (2022). Across all seasons, the trials were planted in mid‐October and harvested until the beginning of March. Strawberry fruits were harvested on a weekly basis, and three measurements were collected: the number of the marketable fruit, the number of non‐marketable fruit (“Culls”), and the total weight (grams) of the marketable fruit (“Marketable Yield”). The data were summed across the growing season to create a single measurement for each plot. Average Weight was calculated as the Marketable Yield divided by the sum of the number of marketable fruits harvested. In addition to these traits, soluble solid content (Brix) and Firmness were measured at three, four, or five separate weeks depending on the growing season. A handheld digital refractometer was used on a single fruit from each plot. Firmness was evaluated by manually squeezing an individual strawberry and rating it on a 1–5 scale.

The yield trials were spatially corrected using the “SpATS” R package (Velazco et al., 2017) with the genotype and experimental bed set as fixed effects. The row and column effects were solved as random effects. The “SpATS” function using default parameters was solved, and plots with a residual >3.5 standard deviations were removed from the analysis. Due to inconsistency of fruit within each week, which resulted in more sparse data, Brix and Firmness were not spatially corrected.

To account for season effects, replication effects, and genotype‐by‐environment interactions, a mixed model was used for Average Weight, Culls, and Marketable Yield. The mixed models were solved using the “lmer” function in the “lme4” R package (Bates et al., 2015). The following equation is used:

$$y_{ijk}=\mathrm{\mu }+s_i+s_i{r}_l+g_k+s_i\times g_k+{\mathrm{\epsilon }}_{ijk}$$ (1)

where y is the phenotypic value, $\mathrm{\mu }$ is the overall mean, s is the fixed effect of the growing season, r is the fixed effect of the replication nested within each season, g is the random effect of the genotype, s × g is the random interaction effect of season‐by‐genotype interaction, and $\mathrm{\epsilon }$ is the residual. Hence, $y_{ijk}$ would be the phenotypic value of the plot in the ith season, the jth replication, and the kth genotype. A separate mixed model was used for Brix and Firmness that included the week of measurement as a fixed effect defined as t.

$$y_{ijkl}=\mathrm{\mu }+s_i+s_i{t}_j+s_i{r}_k+g_l+s_i\times g_l+{\mathrm{\epsilon }}_{ijkl}$$ (2)

All other variables in Equation (2) were previously defined in Equation (1) and the week of measurement was nested within each season. Broad‐sense heritability was calculated as $H^2={\mathrm{\sigma }}_G^2/({\mathrm{\sigma }}_G^2+{\mathrm{\sigma }}_{G\times S}^2+{\mathrm{\sigma }}_R^2)$, where ${\mathrm{\sigma }}_G^2$ was the genetic variance, ${\mathrm{\sigma }}_{G\times S}^2$ was the genotype‐by‐season variance, and ${\mathrm{\sigma }}_R^2$ was the residual variance. These variance components were solved using Equations (1) and (2) for all phenotypes. A likelihood ratio test was used to test the significance of the genetic variance.

2.3. Phenotypic analyses for pathology trials

Four different strawberry diseases were evaluated in this study: phytophthora crown rot (PhCR) caused by the pathogen Phytophthora cactorum, charcoal rot (CR) caused by the pathogen Macrophomina phaseolina, colletotrichum crown rot (CCR) caused by the Colletotrichum gloeosporioides species complex, and powdery mildew (PM) caused by the pathogen Podosphaera aphanis. For all four disease phenotypes, the resistance level was measured in inoculated studies, and the genotypic effects were estimated using Equation (1). Broad‐sense heritability was calculated using the method previously described.

A total of 3819 unique genotypes were evaluated in inoculated PhCR trials grown across eight seasons during the years 2013–2021. The inoculated genotypes were a mixture of seedlings and advanced clones derived from 564 unique biparental families with a mean family size of five. For a given genotype, the median number of full‐sibs was seven and the median number of half‐sibs was 67 within a season, whereas the median number of full‐sibs was zero and the median number of half‐sibs was 92 across seasons. A more detailed description of the inoculation protocol and procedures can be found in Mangandi et al. (2017). In all eight seasons, a disease trial was conducted using a 10‐plant plot in a two‐replication RCB experiment. In addition to these trials, a three‐replication RCB experiment using a single‐plant plot was phenotyped during the 2013–2014 and the 2014–2015 seasons. Data were scored as plant deaths on a weekly basis for up to 20 weeks after planting. The weekly plant death scores were used to calculate the area under the disease progress curve (AUDPC) for each plot (Campbell & Madden, 1990). AUDPC measurements were calculated on a per‐plant basis, and then an average per plot was used to standardize the two different plot types.

A total of 1938 unique genotypes were evaluated in inoculated CR trials across six seasons between the years 2017–2023. The inoculated genotypes were a mixture of seedlings and advanced clones derived from 246 unique biparental families with a mean family size of five. For a given genotype, the median number of full‐sibs was 18 and the median number of half‐sibs was 50 within a season, whereas the median number of full‐sibs was zero and the median number of half‐sibs was 44 across seasons. A more detailed description of the inoculation protocol and procedures can be found in Nelson et al. (2021). All CR trials were designed as a two‐replication RCB experiment using four‐plant plots. Data were scored as collapsed plants on a weekly basis and used to calculate AUDPC for each plot.

A total of 1517 unique genotypes were evaluated in inoculated CCR trials across eight seasons between the years 2013 and 2017 and 2018 and 2022. The inoculated genotypes were a mixture of seedlings and advanced clones derived from 298 unique biparental families with a mean family size of four. For a given genotype, the median number of full‐sibs was seven and the median number of half‐sibs was 38 within a season, whereas the median number of full‐sibs was zero and the median number of half‐sibs was 17 across seasons. A more detailed description of the inoculation protocol and procedures can be found in Anciro et al. (2018). For the season 2013–2014, CCR trials were evaluated using a three‐replication RCB design where each plot was represented by a single plant. In all other seasons, CCR trials were evaluated using a two‐replication RCB design with plots containing four or eight clonal replicates. Data were scored as collapsed plants on a weekly basis and used to calculate AUDPC for each plot. AUDPC measurements were calculated on a per‐plant basis, and then an average per plot was used to standardize the two different plot types.

A total of 1829 unique genotypes were evaluated for natural infection of PM across five seasons between 2017 and 2021 and in 2022 and 2023. The rated genotypes were primarily represented by seedlings derived from 26 biparental families with a mean family size of 65. For a given genotype, the median number of full‐sibs was 73 and the median number of half‐sibs was 304 within a season, whereas the median number of full‐sibs was zero and the median number of half‐sibs was 202 across seasons. Other than these families, there were 35 genotypes in this study that were parents and check varieties. PM was scored on a 0–6 scale at a single timepoint. More methodology and population details can be found in Tapia et al. (2022). In all seasons, trials were evaluated as a three‐replication RCB design where plots were represented by single plants.

2.4. Genotypic data and genetic map

The individuals represented in this dataset were genotyped using two separate genotyping arrays: the iStraw 35k single‐nucleotide polymorphism (SNP) array (van de Weg et al., 2016) or the Fana 50k SNP array (Hardigan, Feldmann, et al., 2021). There were 3218 individuals genotyped on the 50k SNP array and 3651 individuals genotyped on the 35k SNP array. There were 151 individuals genotyped using both SNP arrays. A set of 4242 overlapping SNPs was used to align the two sets of genotypes and to impute the 35k genotypes onto the 50k array. A more detailed description of the imputation process can be found in Sleper et al. (2025). The average imputation accuracy was 95% across the entire genome while subgenome imputation accuracy was highly consistent (A = 96%, B = 95%, C = 94%, D = 95%). A total of 42,543 SNPs were used in this study that could be definitively mapped to each chromosome according to Jiménez et al. (2023). The subgenomic distribution of SNPs was mostly even (A = 11,366, B = 10,892, C = 10,255, D = 10,030) and aligned with the subgenome nomenclature from Session and Rokhsar (2023).

2.5. Additive genome model

To test the importance of modeling subgenome dominance and epistatic effects in allopolyploid strawberries, we sequentially tested models represented in Table 2.

TABLE 2.

Description and equations of genomic selection models.

Model	Equation	Kernels
Additive	$y=1\mathrm{\mu }+Zg+e$	Additive
Subgenome	$y=1\mathrm{\mu }+Z{g}_A+Z{g}_B+Z{g}_C+Z{g}_D+e$	Subgenome additive
Additive+Epistasis	$y=1\mathrm{\mu }+Zg+Z{g}_E+e$	Additive and Epistatic
Subgenome+Epistasis	$y=1\mathrm{\mu }+Z{g}_A+Z{g}_B+Z{g}_C+Z{g}_D+Z{g}_{A\#B}+Z{g}_{A\#C}+Z{g}_{A\#D}+Z{g}_{B\#C}+Z{g}_{B\#D}+Z{g}_{C\#D}+e$	Subgenome additive and subgenome epistatic
Subgenome+Covariance	$y=1\mathrm{\mu }+Z{g}_A+Z{g}_B+Z{g}_C+Z{g}_D+Z{g}_{AB}+Z{g}_{AC}+Z{g}_{AD}+Z{g}_{BC}+Z{g}_{BD}+Z{g}_{CD}+e$	Subgenome additive and Subgenome covariance

As a benchmark, we relied on a traditional Additive model (“Additive”), represented by the following equation:

$$\mathit{y}=1\mathrm{\mu }+\mathit{Zg}+\mathit{e}$$ (3)

where $\mathit{y}$ is the vector of genetic value from the previous phenotypic correction models (Equations 1 and 2), $1$ is a vector of ones with the same dimension as $\mathit{y}$, $\mathrm{\mu }$ is the population mean, $\mathit{g}$ is the vector of additive genetic effects, $\mathit{Z}$ is the incidence matrix for the genetic random effects, and $\mathit{e}$ is the vector of residuals. A flat prior was assigned to the mean ($\mathrm{\mu }),$ following a Gaussian distribution with a mean of zero and a variance of ${10}^{10}$. The vectors $\mathit{g}$ and $\mathit{e}$ were assumed to follow Gaussian distributions: $\mathit{g}\sim N(0,\mathit{G}{\mathrm{\sigma }}_g^2)$ and $\mathit{e}\sim N(0,\mathit{I}{\mathrm{\sigma }}_e^2)$, where $\mathit{I}$ is the identity matrix, ${\mathrm{\sigma }}_g^2$ represents the additive genetic variance, and ${\mathrm{\sigma }}_e^2$ is the residual variance. $\mathit{G}$ is the additive genomic relationship matrix, constructed as the realized genetic relationship proposed by VanRaden (2008) and implemented in the “AGHmatrix” R package (Amadeu et al., 2016).

2.6. Additive subgenome model

The genetic value ($\mathit{g})$ in Equation (3) can be decomposed into independent additive effects for each of the four strawberry subgenomes, that is, $\mathit{g}={\mathit{g}}_{\mathit{A}}+{\mathit{g}}_{\mathit{B}}+{\mathit{g}}_{\mathit{C}}+{\mathit{g}}_{\mathit{D}}$. Instead of using a single kernel associated with the genetic values in the model, four separate kernels were included, each representing a subgenome. This model will be referred to as the “Subgenome” model hereafter. This approach allows us to evaluate the contribution of each subgenome to phenotypic variation and is represented by:

$$\mathit{y}=1\mathrm{\mu }+\mathit{Z}{\mathit{g}}_{\mathit{A}}+\mathit{Z}{\mathit{g}}_{\mathit{B}}+\mathit{Z}{\mathit{g}}_{\mathit{C}}+\mathit{Z}{\mathit{g}}_{\mathit{D}}+\mathit{e}$$ (4)

In Equation (4), each subgenome has its own genetic value, genetic variance, and relationship between individuals. Thus, ${\mathit{g}}_{\mathit{i}}$ is the vector of additive genetic effects for the ith subgenome ($i=A,B,C,D)$, with ${\mathit{g}}_i\sim N(0,{\mathit{G}}_{\mathit{i}}{\mathrm{\sigma }}_{g_i}^2).$ Here, ${\mathrm{\sigma }}_{g_i}^2$ denotes the additive genetic variance associated with each subgenome, and ${\mathit{G}}_{\mathit{i}}$ is the additive subgenome relationship matrix.

We also tested a unique case of the Subgenome model with only a single subgenome kernel. In this case, only the ith subgenome was used in the model; therefore, four different iterations were run, testing each subgenome independently.

2.7. Epistasis model

A genomic model that includes additive and epistatic genetic effects (“Additive+Epistasis”) was used (Henderson, 1985). Epistasis was calculated as the Hadamard product of the genomic relationship matrices. An additional kernel is added to Equation (3) to account for an additive‐by‐additive epistatic interaction term, as follows:

$$\mathit{y}=1\mathrm{\mu }+\mathit{Zg}+\mathit{Z}{\mathit{g}}_E+\mathit{e}$$ (5)

where ${\mathit{g}}_{\mathit{E}}\sim N(0,{\mathit{G}}_E{\mathrm{\sigma }}_{g_E}^2)$, ${\mathit{G}}_E=\mathit{G}\#\mathit{G}$, $\#$ is the Hadamard product operator, and ${\mathrm{\sigma }}_{g_E}^2$ is the genetic variance associated with the epistatic effects.

2.8. Subgenome epistasis model

Similarly to the additive effect, the epistatic interaction effect can also be decomposed into interactions across subgenomes, that is, ${\mathit{g}}_E={\sum }_{\mathit{i}}{\sum }_{\mathit{j}}{\mathit{g}}_{\mathit{i}\#\mathit{j}}$ ($i,j=A,B,C,D$ and $i\ne j$), since there are no markers in common across subgenomes (Martini et al., 2016; Santantonio et al., 2019). Thus, Equation (5) becomes:

$$\begin{array}{ccc}\hfill \mathit{y}& =& 1\mathrm{\mu }+\mathit{Z}{\mathit{g}}_{\mathit{A}}+\mathit{Z}{\mathit{g}}_{\mathit{B}}+\mathit{Z}{\mathit{g}}_{\mathit{C}}+\mathit{Z}{\mathit{g}}_{\mathit{D}}+\mathit{Z}{\mathit{g}}_{\mathit{A}\#\mathit{B}}+\mathit{Z}{\mathit{g}}_{\mathit{A}\#\mathit{C}}\hfill \\ & & +\mathit{Z}{\mathit{g}}_{\mathit{A}\#\mathit{D}}+\mathit{Z}{\mathit{g}}_{\mathit{B}\#\mathit{C}}+\mathit{Z}{\mathit{g}}_{\mathit{B}\#\mathit{D}}+\mathit{Z}{\mathit{g}}_{\mathit{C}\#\mathit{D}}+\mathit{e}\hfill \end{array}$$ (6)

where ${\mathit{g}}_{\mathit{i}\#\mathit{j}}\sim N(0,{\mathit{G}}_{i\#j}{\mathrm{\sigma }}_{g_{i\#j}}^2)$, ${\mathit{G}}_{i\#j}={\mathit{G}}_{\mathit{i}}\#{\mathit{G}}_{\mathit{j}}$ and ${\mathrm{\sigma }}_{g_{i\#j}}^2$ is the genetic variance associated with the epistatic effects between the ith and jth subgenomes. This model will be referred to as the “Subgenome+Epistasis” model.

2.9. Across subgenome epistasis model

Under Hardy‐Weinberg equilibrium, subgenomes segregate independently; consequently, subgenome genetic effects do not have covariance between them, that is, $Cov({\mathit{g}}_{\mathit{i}},{\mathit{g}}_{\mathit{j}})=0$ where $i,j=A,B,C,D$ with $i\ne j$. However, in a breeding program under intense selection pressure, this assumption does not hold. Therefore, we constructed a model that allows for covariance between subgenomes that may be present in a breeding program and is inspired by the transcriptome study of Zhou et al. (2020), which is as follows:

$$\begin{array}{ccc}\hfill \mathit{y}& =& 1\mathrm{\mu }+\mathit{Z}{\mathit{g}}_{\mathit{A}}+\mathit{Z}{\mathit{g}}_{\mathit{B}}+\mathit{Z}{\mathit{g}}_{\mathit{C}}+\mathit{Z}{\mathit{g}}_{\mathit{D}}+\mathit{Z}{\mathit{g}}_{\mathit{AB}}+\mathit{Z}{\mathit{g}}_{\mathit{AC}}\hfill \\ & & +\mathit{Z}{\mathit{g}}_{\mathit{AD}}+\mathit{Z}{\mathit{g}}_{\mathit{BC}}+\mathit{Z}{\mathit{g}}_{\mathit{BD}}+\mathit{Z}{\mathit{g}}_{\mathit{CD}}+\mathit{e}\hfill \end{array}$$ (7)

where ${\mathit{g}}_{\mathit{ij}}\sim N(0,{\mathit{G}}_{ij}{\mathrm{\sigma }}_{\mathit{ij}})$, , $\sqrt{{\mathit{G}}_{\mathit{i}}}$ is the Cholesky decomposition of the genetic relationship kernel matrix with and ${\mathrm{\sigma }}_{\mathit{ij}}$ is the genetic covariance between the ith and jth subgenomes. This model will be referred to as the “Subgenome+Covariance” model.

2.10. Genetic parameter estimation

The Bayesian genomic prediction models were implemented using the “BGLR” package in R software (Pérez & De Los Campos, 2014), with 500,000 iterations for the Markov chain Monte Carlo (MCMC) algorithms, a burn‐in period of 50,000 MCMC cycles, and a thinning rate of 10 before saving samples, resulting in a total of 45,000 MCMC cycles. The genetic vectors in the previous models, generally denoted by $\mathit{g}$, are assumed to follow a normal distribution with zero mean e covariance matrix $\mathit{G}{\mathrm{\sigma }}_g^2$ ($N(0,\mathit{G}{\mathrm{\sigma }}_g^2)$) (i.e., where $\mathit{G}$ is the relationship matrix associated to the specific genetic effect and its respective genetic variance component. The mean $\mathrm{\mu }$ is assumed to follow a flat prior, $N(0,{10}^{10})$, while the genetic and residual variance components are assumed a scaled‐inverse chi‐square distribution. All hyperparameters were defined following Pérez and de los Campos (2014). Autocorrelations and convergence tests were performed using the Geweke criteria (Geweke, 1992) for the residual and genetic chains of the model, respectively. For each model, model fit was calculated as the deviance information criterion (DIC).

For comparing estimates of genetic variances across different models, the procedure described by Legarra (2016) was followed. All genetic variance components were corrected using $s_g^2=(\overline{diag(\mathit{G})}-\overline{\mathbf{G}}){\mathrm{\sigma }}_g^2$. The contribution of each subgenome was estimated using the Subgenome model as $c_i^2=\frac{s_{g_i}^2}{s_g^2}$, where $s_g^2$ is the corrected additive genetic component and $s_{g_i}^2$ is the corrected additive genetic variance for the ith subgenome for $i=A,B,C,D$. We assessed the contribution of the different genetic architecture components using the Additive+Epistasis, Subgenome+Epistasis, and Subgenome+Covariance models. Using the Additive+Epistasis model, epistasis contribution was calculated using $c_E^2=\frac{s_{g_E}^2}{s_g^2+s_{g_E}^2}$, where $s_g^2$ was the corrected additive genetic component and $s_{gE}^2$ was the corrected epistatic genetic component. Next, using the Subgenome+Epistasis model the contribution of epistasis between subgenomes was computed as the following $c_{i\#j}^2=\frac{s_{g_{i\#j}}^2}{{\sum }_i{s}_{g_i}^2+{\sum }_i{\sum }_{\begin{array}{c}j\\ i\ne j\end{array}}{s}_{g_{i\#j}}^2}$, where $s_{g_{i\#j}}^2$ is the corrected epistasis between the ith and the jth subgenomes, ${\sum }_i{s}_{g_i}^2$ is the corrected additive genetic variance for each subgenome, and ${\sum }_i{\sum }_{\begin{array}{c}j\\ i\ne j\end{array}}{s}_{g_{i\#j}}^2$ is the corrected epistasis between all other subgenome contributions that are not specifically the ith and the jth subgenomes. Lastly, using the Subgenome+Covariance model, the contribution of covariance between subgenomes was calculated as $c_{ij}^2=\frac{s_{g_{ij}}^2}{{\sum }_i{s}_{g_i}^2+{\sum }_i{\sum }_{\begin{array}{c}j\\ i\ne j\end{array}}{s}_{g_{ij}}^2}$, where the individual components were identical to the genetic contributions of the Subgenome+Epistasis model, but the subgenome epistasis is replaced with subgenome covariance. Additionally, we constructed highest density probabilities (HPD) intervals to compare the ratio of variance components estimated within each model. In these HPD intervals, if the interval included one, then the variances of the two genetic effects were deemed equal; if the interval values were all greater than one, the first variance component was considered larger than the second; and if the interval values were all less than one, the opposite was true.

2.11. Model comparison

To test the importance of accounting for subgenome dominance and epistatic effects in genomic selection for allopolyploid strawberries, we measured the predictive performance of each model using cross‐validation. To this end, a delete‐season scheme was implemented, where, systematically, a single season was removed from the training data set and considered as the validation set. The training set for the yield trials consisted of nine seasons, and for the pathology trials, depending on the trait, consisted of five to seven seasons. Overlapping genotypes were also removed from the validation set. The predictions presented in this analysis are of the category “predicting new genotypes in a new environment.” This cross‐validation method was chosen because it best represents how genomic selection would be used within a breeding program and is the most rigorous test for new models. An additional cross‐validation scenario was run where genotypes across all seasons were included in a 10‐fold cross‐validation scheme. In this scenario, the genetic values were calculated for each trait across all seasons using Equations (1) and (2). Each 10‐fold cross‐validation was performed five times. This cross‐validation scenario was chosen to reduce the effect of genotype‐by‐environment interaction on the test set and to include more highly related individuals between the training sets and the test sets (i.e., full‐sibs). Since individual seasons generally represented a new breeding cycle in our dataset, the delete‐season cross‐validation scheme eliminated most full‐sib relationships from the training datasets, but using the 10‐fold cross‐validation, a higher degree of genetic relatedness was more likely to occur between training and test sets. Predictive abilities were calculated as the Pearson correlation between the predicted genetic effects from each genomic selection model and the observed genetic effects in the validation season based on Equations (1) and (2).

3. RESULTS

Across the seasons, the estimated broad‐sense heritability values for the nine phenotypes ranged from low (firmness = 0.13) to high (PM = 0.8) (Table 3). Estimates for genetic variance were significant (p = 0.05) for all nine phenotypes. When compared, the genotype‐by‐season effect relative to genetic variance ranged widely across the traits, with Brix, PhCR, and PM demonstrating low levels (6%–12%). In contrast, Average Weight (59%), Culls (40%), Firmness (39%), Marketable Yield (68%), CR (59%), and CCR (68%) demonstrated higher levels of genotype‐by‐season variation relative to genetic variance.

TABLE 3.

Variance component estimates for nine strawberry phenotypes, where V _G is the genetic variance, V _GxS is the genotype‐by‐season variance, V _R is the residual variance, and H ² is the broad‐sense heritability.

Phenotype	V G	V GxS	V R	H 2
Average weight	0.61	0.36	0.02	0.61
Brix	0.17	0.01	0.82	0.17
Culls	0.70	0.28	0.02	0.70
Firmness	0.12	0.05	0.83	0.13
Marketable yield	0.58	0.39	0.03	0.58
PhCR	0.41	0.04	0.55	0.41
CR	0.55	0.33	0.12	0.55
CCR	0.47	0.32	0.21	0.47
PM	0.80	0.10	0.10	0.80

Note: All variance components are presented as the proportion of the total variation and all estimates of genetic variance were significant (p = 0.05).

Abbreviations: CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

3.1. Subgenome effects

The level of genetic contribution for each subgenome varied widely among phenotypes (Figure 1). On average across all nine phenotypes, subgenome B contributed the most genetic variation at 33%, with subgenome A next at 27%, then subgenome C at 23%, and finally subgenome D at 17%. The contributions of subgenomes A (range = 7.7%–55.8%) and B (range = 13.3%–68.2%) varied across phenotypes considerably more than subgenomes C (range = 14.1%–32.9%) and D (range = 13.3%–21.3%). Also relevant, the subgenome contribution patterns were highly variable among phenotypes. For example, for Average Weight, Brix, Culls, Marketable Yield, and PM, we noticed a balanced contribution across subgenomes with a maximum rate between 28.1% and 33.8%. On the other hand, Firmness, PhCR, CR, and CCR demonstrated unbalanced subgenome contribution patterns with maximum contribution rates between 51.1% and 68.2%.

FIGURE 1.

Additive genetic contributions of each subgenome to each phenotype using the subgenome genomic selection model. CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

3.2. Genetic architecture

For all nine phenotypes, additive genetic variance accounted for a majority of the total genetic variance (Figure 2). Based on the Additive+Epistasis model, epistasis accounted for an average of 28% of the genetic variation, with as little as 12% for Average Weight and as much as 46% for Culls. When incorporating subgenome effects and subgenome interactions using the Subgenome+Epistasis model, the proportion of additive genetics remained mostly unchanged except for CR, which decreased from 79% to 65%. Using HPD intervals for the Subgenome+Epistasis model, there was no single interaction between subgenomes that was substantially more contributive to the total genetic variation than other subgenome interactions. Across all phenotypes, the genetic contribution for the six different subgenome interactions was mostly balanced. Similar results were observed for the Subgenome+Covariance model. In this case, similar levels of epistasis were observed as for the Subgenome+Epistasis model except for Culls, which decreased in epistatic variance from 50% to 21%. Again, similar to the Subgenome+Epistasis model, the genetic contributions were mostly balanced for the six different subgenome covariance effects; however, HPD intervals revealed that 11 out of the 135 subgenome covariance effect comparisons across the nine phenotypes were different. These subgenome interactions were BD < CD for Marketable Yield, Brix, PM, and CR; AD < BC for PhCR, PM, and CR; AD < CD for PhCR, PM, and CR; and AD < BC for CCR.

FIGURE 2.

Sources of genetic variation for nine strawberry phenotypes using Additive+Epistasis, Subgenome+Epistasis, and Subgenome+Covariance genomic selection models. CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

3.3. Model comparison using cross‐validation

For the delete‐season cross‐validation across all genomic selection models and traits, average predictive abilities were moderate to high (range = 0.37–0.68) (Figure 3). Based on a Tukey's test, none of the genomic selection models tested were significantly better than each other (p = 0.05). Despite the lack of significance, the Additive+Epistasis, Subgenome, and Subgenome+Epistasis models consistently had equivalent or higher predictive ability than the baseline Additive model across all nine phenotypes. Relative to the Additive model, the predictive ability for the Subgenome+Epistasis model was 3% better on average than the Additive model. Additionally, the predictive ability for the Subgenome+Epistasis model was 6% better for Marketable Yield and 7% better for CCR. Even with no epistasis term, the predictive ability for the Subgenome model was on average 2% greater than the Additive model across all phenotypes. Relative to the Additive model, the predictive ability for the Subgenome+Covariance model was 1% better than the Additive model; however, this was inconsistent across phenotypes, and a decrease in predictive ability of 1% was observed for Brix and PM.

FIGURE 3.

Average predictive ability of five different genomic selection models using a delete‐season cross‐validation scheme across nine strawberry phenotypes. None of the predictive abilities were more significant than the others (p = 0.05) using a Tukey's test. CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

Predictive abilities were higher in the 10‐fold cross‐validation than the delete‐season cross‐validation, except for Average Weight and Brix. Based on a Tukey's test, the Subgenome+Epistasis model was significantly better than all other models for Firmness, Marketable Yield, PhCR, CR, and PM (p = 0.05) (Figure 4). For Average Weight, Brix, Culls, and CCR, the Additive+Epistasis and the Subgenome+Epistasis models were significantly better than the Additive, Subgenome, and Subgenome+Covariance models (p = 0.05). In this scenario, the Subgenome+Epistasis model improved predictive ability by 8%–34% relative to the Additive model. Additionally, the Additive+Epistasis model improved predictive ability by 6%–31% relative to the Additive model. No differences in predictive ability were observed among the Additive, Subgenome, and Subgenome+Covariance models.

FIGURE 4.

Average predictive abilities for five different genomic selection models using a 10‐fold cross‐validation. Mean separation letters represent significant differences for Tukey's tests (p = 0.05). CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

Substantial decreases in predictive ability were observed for the Subgenome model with a single subgenome kernel compared to the Subgenome model with all four subgenomes (Figure 5). The full Subgenome model was significantly better than models incorporating a single subgenome for Average Weight, Brix, Firmness, and PM (p = 0.05). Across all phenotypes, the average predictive ability decreased by 15%–48% for subgenomes exhibiting the lowest predictive ability. On the other hand, the average predictive ability across all phenotypes only decreased by 7%–19% for subgenomes demonstrating the highest predictive ability. The most extreme example was observed for PhCR, where predictive ability for subgenome A was 0.26 compared to 0.46 for subgenome B. For Firmness, PhCR, CR, and CCR, which demonstrated unbalanced subgenome contributions, unbalanced subgenome predictive abilities were also observed.

FIGURE 5.

Average predictive abilities for the subgenome genomic selection model with each individual subgenome kernel (A, B, C, and D) and all four subgenome kernels (G). Mean separation letters represent significant differences for Tukey's tests (p = 0.05). CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

3.4. Model comparison using DIC

Across all nine phenotypes, we observed the lowest DIC values for the Subgenome+Epistasis model (Table 4). Additionally, the Additive+Epistasis model was associated with the second lowest DIC for all phenotypes.

TABLE 4.

Deviance information criterion (DIC) for all genomic selection models across nine phenotypes.

Model	Average weight	Brix	Culls	Firmness	Marketable yield	PhCR	CR	CCR	PM
Additive	13,666.3	−1742.0	16,952.7	566.9	9102.1	13,400.6	42,594.0	25,706.2	3240.8
Subgenome	13,675.5	−1739.8	16,948.5	548.6	9074.0	13,378.3	42,583.2	25,630.8	3242.4
Additive+Epistasis	13,481.5	−1995.1	16,572.1	207.4	8997.2	13,317.2	42,360.6	25,479.9	3070.9
Subgenome+Epistasis	13,366.4	−2095.1	16,522.1	59.6	8892.2	13,221.8	42,299.4	25,317.6	3015.0
Subgenome+Covariance	13,675.6	−1725.6	16,977.9	557.1	9082.6	13,401.1	42,587.6	25,641.4	3242.0

Note: Lower values of DIC reflect better model fits, and the best model fit for each trait has been bolded.

Abbreviations: CCR, colletotrichum crown rot; CR, charcoal rot; PhCR, phytophthora crown rot; PM, powdery mildew.

4. DISCUSSION

4.1. Effect of subgenome dominance on phenotype was variable

In this study, subgenome contributions were highly variable among the phenotypes. We observed greater levels of genetic variance associated with subgenome A for Firmness and CR than for other subgenomes. Subgenome B demonstrated the largest contribution to genetic variation for Average Weight, Marketable Yield, CCR, and PhCR. Additionally, subgenome C exhibited the largest contribution to genetic variation for Culls and PM. The only subgenome that did not exhibit the greatest level of genetic variation for any phenotype was subgenome D. Interestingly, subgenome B contributed the most genetic variation across the nine phenotypes. From this vantage point, our results suggest that each subgenome contributes substantially to strawberry phenotypes and that different patterns of genetic control at the subgenome level have arisen for each phenotype. This finding is consistent with a previous wheat study that showed cell type and time‐specific subgenome dominance using transcriptomic data (Pfeifer et al., 2014). In that study subgenome dominance varied widely based on a co‐expression analysis, and each subgenome demonstrated dominance for various gene networks.

The strawberry subgenome that exhibited dominance for a phenotype was sometimes associated with known major loci, but this was inconsistent. For example, subgenome A expressed 51% of the genetic variance for CR, and FaRMp1 is located on chromosome 2A, which is a major effect locus for CR resistance (Nelson et al., 2021). Despite this association, known large effect loci for CR resistance are also found on chromosomes 4B and 4D, which suggests that subgenome A is contributing additional polygenic variation. For PhCR, subgenome B expressed 68% of the genetic variation, which is consistent with the known large effect loci FaRPc2 on chromosome 7B (Mangandi et al., 2017). A separate study using a strawberry population not represented in this study reported a different locus associated with PhCR resistance within subgenome B located on chromosome 6B (Jiménez et al., 2023). For CCR, a known major‐effect locus has been reported by Anciro et al. (2018) on chromosome 6B, which corroborates with high levels of genetic variation explained by subgenome B in this study. For Firmness, a major effect gene, PG1‐6A1, has been mapped and validated (Jiménez et al., 2025); however, this gene has not been previously mapped in the University of Florida strawberry population and is most likely fixed in elite breeding populations, meaning it does not contribute to genetic variation. The other phenotypes in this study were highly polygenic and demonstrated substantially less subgenome dominance. Taken together, these results highlight that underlying genetic loci contributing to various phenotypes in strawberry are spread across all subgenomes.

These results shed new light on the many studies showing dominance of subgenome A in strawberry at the molecular level (Edger et al., 2019; Fang et al., 2024; Han et al., 2025; Song et al., 2024). Additionally, our study is fundamentally different than prior strawberry subgenome dominance investigations. First, our study was not focused on molecular data from a single strawberry variety but on phenotypic data collected over an entire decade. Second, these results represent not the entire species but a single strawberry breeding program, in which many loci are fixed and not segregating, and thus would not contribute to genetic variance. It is also important to understand that subgenome A dominance at the molecular level is subtle. For example, in a recent study, subgenome A contained 1913 more genes than subgenome B, 2367 more genes than subgenome C, and 3541 more genes than subgenome D (Han et al., 2025). Taken as a whole, this means that the other three subgenomes have gene content between 85% and 93% of subgenome A and that subgenome A accounts for 27% of the total gene content. These gene content proportions are directly relevant to the breeding population reported here, as Han et al. (2025) used the University of Florida cultivar “Florida Brilliance.” which was included in this study (in addition to many progeny and grand progeny). Altogether, our results suggest that subgenome dominance in strawberry is highly complex and phenotype‐specific, and that our understanding of this phenomenon is in an early stage.

4.2. Genetic architecture of strawberry phenotypes

Our results also suggest that epistasis plays a substantial role in strawberry. This is consistent with previous genomic selection studies in strawberry for Average Weight, Brix, Culls, and Marketable Yield (Zingaretti et al., 2020). Similar reports using genomic selection to estimate epistasis have been recorded in sweet cherry (Piaskowski et al., 2018) and in autopolyploid potato (Endelman et al., 2018). In both crops, epistasis was significant and varied substantially by the phenotype. This is the first time that epistasis models have been applied to the strawberry disease phenotypes used in this study. Additionally, epistasis between subgenomes was observed for all phenotypes, but highly influential subgenome‐by‐subgenome interactions were not detected. Despite this, additive genetic variance accounted for the majority of genetic variance for all phenotypes, which was previously reported (Zingaretti et al., 2020). Due to mathematical complexity, we want to highlight that understanding epistasis and particularly subgenome epistasis requires very large datasets. Despite this being the largest dataset reported on in strawberry genetics (N = 6718), it is still quite small when considering that the strawberry genome contains over 100,000 genes (Edger et al., 2019). Most phenotypes of interest to strawberry breeders are highly polygenic; hence, polygenic epistasis quickly requires very large datasets and extensive computation (Fu et al., 2023). With reduced costs of genotyping and optimization of strawberry trialing (Sleper et al., 2025), we envision datasets like this one, but much larger, being compiled in the near future that should be more sufficient for deeper explorations into polygenic epistasis.

4.3. Genomic selection in strawberry

The largest genomic selection training set used in this study was nearly three times as large as previously reported by the University of Florida strawberry research team (Osorio et al., 2021). With this improvement, we observed an average increase in predictive ability of 39% for the four phenotypes in common between both studies. Additionally, we reported reliable genomic selection predictive abilities for PhCR, CR, and CCR, which had previously not been reported from this strawberry breeding program. These phenotypes have been targeted for marker‐assisted selection due to the presence of large effect loci. This study demonstrated that more genetic variation can be explained using genomewide markers and that genomic selection would provide higher genetic gain than marker‐assisted selection.

Genomic selection models incorporating epistasis and subgenome effects consistently outperformed the baseline Additive model. This was especially true in the 10‐fold cross‐validation scenario where models incorporating epistasis (Additive+Epistasis and Subgenome+Epistasis) were significantly (p = 0.05) better than models with only additive effects. The degree of increased predictive ability for the Subgenome+Epistasis model was in direct correlation (r = 0.89) with the percentage of epistasis estimated by the genomic selection models for each phenotype. Overall, this result demonstrates that estimates of epistasis were not arbitrary and resulted in large increases in effectiveness of genomic selection. Despite the lack of statistical significance in the delete‐season cross‐validation, improved predictive ability was consistent across all nine phenotypes and across a 10‐season dataset. Furthermore, DIC values were lowest for the Subgenome+Epistasis and Additive+Epistasis models, indicating a better explanation of the underlying genetic architecture. Additionally, the delete‐season method of cross‐validation was highly rigorous (prediction of new genotypes in new seasons), suggesting that these improvements were real and could be sustained in future breeding cycles. While the differences were not large (predictive abilities improved relatively 1%–7% depending on phenotype and model), they could compound over multiple breeding cycles to provide meaningful improvements.

Two different cross‐validation methods were incorporated into this analysis to represent prediction across cycles in a recurrent selection breeding program (i.e., predicting a new season) and prediction across all data within the breeding program (i.e., incorporating the new season of data with historical data). The results across these two methods were consistent, but the results were much more stark in the 10‐fold cross‐validation and resulted in significant (p = 0.05) differences for all nine phenotypes. A similar result was obtained when incorporating epistatic effects in genomic selection models in polyploid wheat (Raffo et al., 2022). The authors observed that within each breeding cycle there was a higher degree of relatedness than across breeding cycles, especially full‐sibs. This resulted in a greater predictive ability for epistatic genomic selection models within breeding cycles compared to across breeding cycles. Epistasis should be more likely to be detected in relationships with high levels of genetic similarity (i.e., half‐sibs and full‐sibs) since alleles have higher likelihood of being inherited in tandem due to shared parentage (Raffo et al., 2022). In our study, depending on the phenotype, the median number of full‐sibs increased from zero to 73, and the median number of half‐sibs increased by around 50% when genotypes within a cycle were included in the training set. We suspect that this increase in highly related genotypes was the primary driver behind increased predictive ability for the Additive+Epistasis and Subgenome+Epistasis models.

The inclusion of epistasis terms in genomic selection models is highly important for crops exhibiting large proportions of epistasis. First, for a purely additive genomic selection model, epistatic genetic variation and additive genetic variance are confounded, but in genomic selection models including epistatic terms, these genetic effects can be teased apart. In theory, this would mean that additive genetic variation is more accurately described by these models. Second, parents could be selected that had potential for complementary epistatic effects. While epistatic genetic variation is not inherited per se, accumulating alleles in positive epistatic interactions would be highly beneficial to plant breeding programs. Utilizing epistatic genomic selection models could favor the accumulation of these interactions over multiple breeding cycles. Last, plant breeders of clonally propagated plants like strawberry are interested in selection of the best new clone, not just population improvement. If models using epistatic terms are significantly better (i.e., DIC and predictive ability), then these models would be more effective at estimating the clonal value and not simply the additive term, which is an incomplete expression of total genetic variance. Empirical experiments need to be performed that use these different genomic selection models for selection across multiple cycles to test their true utility (Sleper & Bernardo, 2018). In addition, investigations testing how epistasis affects parent selection methods are needed to inform the selection of cross combinations within a genomic selection framework.

5. CONCLUSIONS

Our study sheds new light on the nature of subgenome dominance in allo‐octoploid strawberry. Despite the molecular dominance of subgenome A, we observed highly variable subgenome contribution levels across phenotypes. Yet subgenome D clearly exhibited the least genetic variation across the traits, in keeping with D having the lowest gene content. We also found evidence that epistasis makes considerable contributions to the complex genetic architecture of nine strawberry phenotypes. Lastly, our study highlights the increasing utility of genomic selection within strawberry breeding programs. As training sets continue to grow, genomic selection predictive abilities continue to increase, and models allowing for complex interactions (i.e., epistasis) are increasingly viable. Continued empirical work is needed to better understand genomic selection models incorporating subgenome effects and epistasis.

AUTHOR CONTRIBUTIONS

Joshua A. Sleper: Data curation; formal analysis; writing—original draft. Camila F. Azevedo: Formal analysis; investigation; methodology; writing—review and editing. Luis F. Ferrao: Conceptualization; methodology; writing—review and editing. Vance M. Whitaker: Funding acquisition; supervision; writing—review and editing.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Sleper, J. A. , Azevedo, C. F. , Ferrao, L. F. , & Whitaker, V. M. (2026). Using genomic selection to examine subgenome dominance and epistasis in allopolyploid strawberry. The Plant Genome, 19, e70204. 10.1002/tpg2.70204

DATA AVAILABILITY STATEMENT

Data used in this experiment may be available upon request.