Overview of an example breeding program

      In a typical strawberry breeding program, the pool of elite parents used to make crosses is updated rapidly according to industry and environmental needs. Because of its octoploid genome, determining the genetic architecture underlying traits in cultivated strawberry is challenging. In addition, strawberry varieties are mostly asexually propagated, and maintaining a strawberry clone from first year field trail to its release is expensive. Every year in the University of Florida (UF) strawberry breeding program, about a hundred crosses are made in the anticipation of superior fruit quality and better resistance to diseases. It takes a year to obtain performance data on progeny and make informed crossesfor the next year. If the genetic potential of future progeny can be estimated prior to field evaluations, increased genetic gains can be realized and resources can be saved.

      An older breeding program like that at UF can utilize years of accumulated phenotypic data and many generations of pedigree information. Most of the elite parents descend from a few common progenitors in the UF strawberry breeding program, which helps magnify pedigree relatedness between training and test populations. In this article, we refer tofour training populations, T1/2013 (Trial 1: year 2013-14), T2/2013 (Trial 2: year 2013-14), T1/2014 (Trial 1: year 2014-15) and T2/2014 (Trial 2: year 2014-15) [3], and a test population: T2/2015 (Trial 2: year 2014-15). In general, T1 trials are of unselected seedlings arising from complex mating designs, and T2 trials are collections of advanced selections that represent the parent pool. Here we utilize this data to demonstrate the ability of FlexQTL™ to predict unknown QTL genotypes and phenotypes for a moderateeffect Soluble Solids Content (SSC) locus on Linkage Group (LG) 6A (unpublished).

      Objectives of this editorial are to (1) describe the ability of FlexQTL™ to predict unknown QTL genotypes and phenotypes using the SSC locus on LG 6A as an example, and (2) demonstrate predictive ability by correlating predicted and observed SSC data for the test population.

Genetic control of a trait and prediction methodology

      The genetic control or architecture of a trait may range from single major locus to a couple of moderate effect loci to many small-effect loci, or any combination thereof. Before predicting unknown phenotypes of individuals, knowledge of the genetic control of a trait is vital. A genome-wide prediction approach is usually effective for a trait control by many small-effect loci as compared to Quantitative Trait Loci (QTL) based MarkerAssisted Selection (MAS) approach with known large-effect marker-trait associations [6]. Several well-established statistical methodologies have been recommended for making genomewide predictions including Genomic Best Linear Unbiased Prediction (GBLUP) and Bayesian methods [3,7]. However, when the genetic architecture of a trait includes at least one identifiable locus, it is valuable to have predictions of QTL alleles, their combinations (genotypes) and their patterns of segregation. In our experience, such a QTL-based approach is particularly effective for a trait controlled primarily by single locus or a couple of loci explaining half or more of the phenotypic variation for the trait.

      The PBA approach using FlexQTL™ software not only detects QTLs but also assigns predicted QTL alleles (QQ, Qq, qq) based on molecular marker allele frequency, Identity By Descent (IBD) probabilities, and available phenotypic data accounting for all known pedigree relationships including immediate parents, grandparents, offspring, siblings and more distant ancestors [4,5,8-10] (Figure 1). FlexQTL™ also helps visualize segregation of QTL genotypes via pedigrees using Pedi map software. Where phenotypic data is missing but marker data is available, phenotypes can be predicted based on the predicted QTL genotypes.

Prediction with FlexQTL™

      The advantages of the PBA approach, both for QTL detection and QTL allele and phenotype prediction, are most apparent in complex population structures [11-19]. Bi-parental experimental designs are inefficient for a large breeding program like the UF strawberry breeding program where the number of parents in a crossing cycle is large and constantly changing. Most QTL analysis methods assume that parents are unrelated, but in reality, most parents in breeding populations are related. FlexQTL™ efficiently utilizes pedigree connections and establishes relatedness among parents in the most complex of population structures [5]. The ability of FlexQTL™ to identify inherited relatedness and re-evaluate pedigree relationships using IBD probability matrices offers superior statistical ability to estimate functional genotypes at a locus and their effects on a trait.

      FlexQTL™ implements phasing of marker alleles based on founder alleles using Linkage Disequilibrium (LD). The tracing of segregating marker alleles from founders to connected parents and progeny generates phased haplotype information for each individual. In addition, where marker information is missing, LDbased estimation is utilized to impute marker alleles [20]. FlexQTL™ efficiently conducts linkage phasing between QTL genotypes (QQ, Qq, qq) and marker genotypes (A/B or A/T/G/C) over diverse genetic backgrounds, providing vital information on the segregation of functional QTL genotypes throughout a breeding population.QTL alleles homozygous for positive effect are represented as QQ [+ +], for heterozygous effects represented as Qq [+ -], and homozygous for negative effects represented as qq [- -]. If an individual’s alleles are not represented sufficiently in the training population, the prediction will be poor. Thus, when separating datasets into training and test populations, relatedness and representation of alleles between the two sets must be carefully considered.

A PBA prediction example

      Soluble solids content is partly controlled by a moderate-effect QTL on LG 6A in the UF breeding program, explaining around 8-15% of phenotypic variation.T1/2013, T2/2013, T1/2014, and T2/2014trials conducted in two different years are considered together as a training population (phenotypic and marker data included in analysis), and T2/2015 is considered as the test population (marker data only included in analysis). The training population includes more than 200 full-sib families with over 1,500 individuals in total. The parents and/or grandparents of the test population individuals are pedigree-connected to the training population. This allows dynamic multi-directional flow of allele information between training and test populations. An example of multi-directional flow of allele information is presented in Figure 1.

      The test populationT2/2015is comprised of approximately 200 selections, the parents of which were represented by numerous half-sibs in the training population. FlexQTL™ simulations were implemented according to [12] using 13 SNP markers spanning the QTL region, and SSC QTL genotypes and phenotypes predicted. In order to estimate Predictive Ability (PA), a Pearson correlation was conducted between observed and predicted SSC data. A positive (r=0.21) and significant (p=0.0027) correlation was observed (Figure 2).

      In addition, Fisher’s Least Significant Difference (LSD) test was conducted to separate observed mean values of each of the predicted QTL genotype classes, and a significant separation (p<0.05) between phenotypic means of each of the QTL genotype classes was observed (Figure 2) [21]. A predictive ability (PA) of 0.21 in genome-wide depends on the context such as population size and diversity, trait heritability and the goal of the study. Given the low to moderate heritability of this trait, we consider 0.21 high enough to warrant further study.

      We believe that as the UF strawberry program accumulates more genotypic and phenotypic data over the years, training populations will grow in size and scope and prediction of unknown QTL genotypes and phenotypes will be more precise. These results hold great value to the UF strawberry breeding program as UF no longer conducts stage-1 replicated trials [3] and information on genetic potential of parents will inform cross combinations. This method is being explored for other traits as well.

      Until now the PBA approach using FlexQTL™ has been employed only for the detection of QTLs and QTL allele analysis for various traits using full phenotypic and marker data [10,22-30]. Here we have explored the ability of FlexQTL™ to predict QTL genotypes and unknown phenotypes where only marker data is available. Preliminary results indicate that it can be a valuable tool to make DNA-informed breeding decisions.

      Based on our experience, several points should be considered:

      1) Traditional genome-wide prediction approaches do not usually isolate the effects of discrete QTL or predict QTL genotypes; however, QTL genotype information from FlexQTL™ can be used as fixed effect for better genome wide predictions using traditional genomic selection models.

      2) Predicted QTL genotypes can be utilized, not only for parent selection to achieve the highest progeny mean, but also to predict phenotypic variances of crosses based on predicted QTL genotype ratios of progenies. If effective, prediction of cross variance would be a unique contribution to the field of genomic prediction.

      3) An important point to mention is that FlexQTL™ requires a genetic or physical map, which is a limitation for some crop species as compared to GBLUP or Bayes B methods, which do not require marker positions.

      4) This methodology will be most advantageous for QTL with moderate effects as opposed to major loci explaining most of the genetic variance for a trait. In the case of a major locus, a single marker in the QTL region is likely to explain the phenotypic variance. In the case of a moderate-effect QTL such as the SSC QTL in this example, FlexQTL™ accounts for the segregation of the 13 SNPs in the QTL region across pedigrees, making QTL predictions that would be almost impossible manually.

      5) Continuing with this line of reasoning, the value of FlexQTL™ should be highest in cases of multiple discrete QTL controlling a trait, which can be simultaneously predicted using this software.

      In summary, FlexQTL™ provides exciting new possibilities for predictive breeding applications that should be especially valuable for traits controlled by one or more moderate-effect QTL.