Biology subjects

Gorjanc, G.

Publications and source records attributed to Gorjanc, G..

At least 55 records · Page 3Linked to original sources

bioRxiv2020

Genomic selection strategies for clonally propagated crops

For genomic selection in clonal breeding programs to be effective, crossing parents should be selected based on genomic predicted cross performance unless dominance is negligible. Genomic prediction of cross performance enables a balanced exploitation of the additive and dominance value simultaneously. Here, we compared different strategies for the implementation of genomic selection in clonal plant breeding programs. We used stochastic simulations to evaluate six combinations of three breeding programs and two parent selection methods. The three breeding programs included i) a breeding program that introduced genomic selection in the first clonal testing stage, and ii) two variations of a two-part breeding program with one and three crossing cycles per year, respectively. The two parent selection methods were i) selection of parents based on genomic estimated breeding values, and ii) selection of parents based on genomic predicted cross performance. Selection of parents based on genomic predicted cross performance produced faster genetic gain than selection of parents based on genomic estimated breeding values because it substantially reduced inbreeding when the dominance degree increased. The two-part breeding programs with one and three crossing cycles per year using genomic prediction of cross performance always produced the most genetic gain unless dominance was negligible. We conclude that i) in clonal breeding programs with genomic selection, parents should be selected based on genomic predicted cross performance, and ii) a two-part breeding program with parent selection based on genomic predicted cross performance to rapidly drive population improvement has great potential to improve breeding clonally propagated crops.

genetics↗

bioRxiv2020

Spatial modelling improves genetic evaluation in smallholder breeding programs

Breeders and geneticists use statistical models for genetic evaluation of animals to separate genetic and environmental effects on phenotype. A common way to separate these effects is to model a descriptor of an environment, a contemporary group or herd, and account for genetic relationship between animals across the environments. However, separating the genetic and environmental effects in smallholder systems is challenging due to small herd sizes and weak genetic connectedness across herds. Our hypothesis was that accounting for spatial relationships between nearby herds can improve genetic evaluation in smallholder systems. Further, geographically referenced environmental covariates are increasingly available and could be used to model underlying sources of the spatial relationships. The objective of this study was therefore to evaluate the potential of spatial modelling to improve genetic evaluation in smallholder systems. We focus solely on dairy cattle smallholder systems. We performed simulations and real dairy cattle data analysis to test our hypothesis. We used a range of models to account for environmental variation by estimating herd and spatial effects. We compared these models using pedigree or genomic data. The results show that in smallholder systems (i) standard models are not able to separate genetic and environmental effects, (ii) spatial modelling increases accuracy of genetic evaluation for phenotyped and non-phenotyped animals, (iii) environmental covariates do not substantially improve accuracy of genetic evaluation beyond simple distance-driven spatial relationships between herds, (iv) the benefit of spatial modelling was the largest when the genetic and environmental effects were hard to separate and (v) spatial modelling was beneficial when using either pedigree or genomic data. We have demonstrated the potential of spatial modelling to improve genetic evaluation in smallholder systems. This improvement is driven by establishing environmental connectedness between herds that enhances separation of the genetic and environmental effects. We suggest routine spatial modelling in genetic evaluations, particularly for smallholder systems. Spatial modelling could also have major impact in studies of human and wild populations.

genetics↗

bioRxiv2020

AlphaPart - R implementation of the method for partitioning genetic trends

BackgroundIn this paper we present the AlphaPart R package, an open-source software that implements a method for partitioning breeding values and genetic trends to identify sources of genetic gain. Breeding programmes improve populations for a set of traits, which can be measured with a genetic trend calculated from averaged year of birth estimated breeding values of selection candidates. While sources of the overall genetic gain are generally known, their realised contributions are hard to quantify in complex breeding programmes. The aim of this paper is to present the AlphaPart R package and demonstrate it with a simulated pig breeding example. ResultsThe package includes the main partitioning function AlphaPart, that partitions the breeding values and genetic trends by analyst defined paths, and a set of functions for handling data and results. The package is freely available from CRAN repository at http://CRAN.R-project.org/package=AlphaPart. We demonstrate the use of the package by examining the genetic gain in a pig breeding example, in which the multiplier achieved higher breeding values than the nucleus for traits measured and selected in the multiplier. The partitioning analysis revealed that these higher values depended on the accuracy and intensity of selection in the multiplier and the extent of gene flow from the nucleus. For traits measured only in the nucleus, the multiplier achieved comparable or smaller genetic gain than the nucleus depending on the amount of gene flow. ConclusionsAlphaPart implements a method for partitioning breeding values and genetic trends and provides a useful tool for quantifying the sources of genetic gain in breeding programmes. The use of AlphaPart will help breeders to better understand or improve their breeding programmes.

genomics↗

bioRxiv2020

Robust Genomic Modelling Using Expert Knowledge about Additive, Dominance and Epistasis Variation

We propose a novel Bayesian approach that robustifies genomic modelling by leveraging expert knowledge through prior distributions. The central component is the hierarchical decomposition of phenotypic variation into additive and non-additive genetic variation, which leads to an intuitive model parameterization that can be visualised as a tree. The edges of the tree represent ratios of variances, for example broad-sense heritability, which are quantities for which expert knowledge is natural to exist. Penalized complexity priors are defined for all edges of the tree in a bottom-up procedure that respects the model structure and incorporates expert knowledge through all levels. We investigate models with different sources of variation and compare the performance of different priors implementing varying amounts of expert knowledge in the context of plant breeding. A simulation study shows that the proposed priors implementing expert knowledge improve the robustness of genomic modelling and the selection of the genetically best individuals in a breeding program. We observe this improvement in both variety selection on genetic values and parent selection on additive values; the variety selection benefited the most. In a real case study expert knowledge increases phenotype prediction accuracy for cases in which the standard maximum likelihood approach did not find optimal estimates for the variance components. Finally, we discuss the importance of expert knowledge priors for genomic modelling and breeding, and point to future research areas of easy-to-use and parsimonious priors in genomic modelling.

genomics↗

bioRxiv2020

Genetics of recombination rate variation in the pig

BackgroundIn this paper, we estimated recombination rate variation within the genome and between individuals in the pig using multiocus iterative peeling for 150,000 pigs across nine genotyped pedigrees. We used this to estimate the heritability of recombination and perform a genome-wide association study of recombination in the pig. ResultsOur results confirmed known features of the pig recombination landscape, including differences in chromosome length, and marked sex differences. The recombination landscape was repeatable between lines, but at the same time, the lines also showed differences in average genome-wide recombination rate. The heritability of genome-wide recombination was low but non-zero (on average 0.07 for females and 0.05 for males). We found three genomic regions associated with recombination rate, one of them harbouring the RNF212 gene, previously associated with recombination rate in several other species. ConclusionOur results from the pig agree with the picture of recombination rate variation in vertebrates, with low but nonzero heritability, and a major locus that is homologous to one detected in several other species. This work also highlights the utility of using large-scale livestock data to understand biological processes.

genetics↗

bioRxiv2020

Phasing and imputation of single nucleotide polymorphism data of missing parents of bi-parental plant populations

This paper presents an extension to a heuristic method for phasing and imputation of genotypes of descendants in bi-parental populations so that it can phase and impute genotypes of parents of bi-parental populations that are fully ungenotyped or partially genotyped. The imputed genotypes of the parent are then used to impute low-density genotyped descendants of the bi-parental population to high-density. The extension works in three steps. First, it identifies whether a parent has no or low-density genotypes available and it identifies all of its relatives that have high-density genotypes. Second, using the high-density information of relatives, it determines whether the parent is homozygous or heterozygous for a given locus. Third, it phases heterozygous positions of the parent by matching haplotypes to its relatives. We implemented the new algorithm in an extension of the AlphaPlantImptue software and tested its accuracy of imputing missing parent genotypes in simulated bi-parental populations from different scenarios. We also tested the accuracy of imputation of the missing parents descendants using the true genotype of the parent and compared this to using the imputed genotypes of the parent. Our results show that across all scenarios, the accuracy of imputation of a parent, measured as the correlation between true and imputed genotypes, was > 0.98 and did not drop below [~] 0.96. The imputation accuracy of a parent was always higher when it was inbred than when it was outbred and when it had low-density genotypes. Including ancestors of the parent at HD, increasing the number of crosses and the number of high-density descendants all increased the accuracy of imputation. The high imputation accuracy achieved for the parent across all scenarios translated to little or no impact on the accuracy of imputation of its descendants at low-density. Key MessageNew fast and accurate method for phasing and imputation of SNP chip genotypes within diploid bi-parental plant populations.

genetics↗

bioRxiv2020

Hierarchical modeling of haplotype effects based on a phylogeny

This paper introduces a hierarchical model to estimate haplotype effects based on phylogenetic relationships between haplotypes and their association with observed phenotypes. In a population there are usually many, but not all possible, distinct haplotypes and few observations per haplotype. Further, haplotype frequencies tend to vary substantially - few haplotypes have high frequency and many haplotypes have low frequency. Such data structure challenge estimation of haplotype effects. However, haplotypes often differ only due to few mutations and leveraging these similarities can improve the estimation of haplotype effects. There is extensive literature on this topic. Here we build on these observations and develop an autoregressive model of order one that hierarchically models haplotype effects by leveraging phylogenetic relationships between the haplotypes described with a directed acyclic graph. The phylogenetic relationships can be either in a form of a tree or a network and we therefore refer to the model as the haplotype network model. The haplotype network model can be included as a component in a phenotype model to estimate associations between haplotypes and phenotypes. The key contribution of this work is that by leveraging the haplotype network structure we obtain a sparse model and by using hierarchical autoregression the flow of information between similar haplotypes is estimated from the data. We show with a simulation study that the hierarchical model can improve estimates of haplotype effects compared to an independent haplotype model, especially when there are few observations for a specific haplotype. We also compared it to a mutation model and observed comparable performance, though the haplotype model has the potential to capture background specific effects. We demonstrate the model with a case study of modeling the effect of mitochondrial haplotypes on milk yield in cattle.

genomics↗

bioRxiv2019

AlphaFamImpute: high accuracy imputation in full-sib families from genotype-by-sequencing data

SummaryAlphaFamImpute is an imputation package for calling, phasing, and imputing genome-wide genotypes in outbred full-sib families from single nucleotide polymorphism (SNP) array and genotype-by-sequencing (GBS) data. GBS data is increasingly being used to genotype individuals, especially when SNP arrays do not exist for a population of interest. Low-coverage GBS produces data with a large number of missing or incorrect naive genotype calls, which can be improved by identifying shared haplotype segments between full-sib individuals. Here we present AlphaFamImpute, an algorithm specifically designed to exploit the genetic structure of full-sib families. It performs imputation using a two-step approach. In the first step it phases and imputes parental genotypes based on the segregation states of their offspring (that is, which pair of parental haplotypes the offspring inherited). In the second step it phases and imputes the offspring genotypes by detecting which haplotype segments the offspring inherited from their parents. With a series of simulations we find that AlphaFamImpute obtains high accuracy genotypes, even when the parents are not genotyped and individuals are sequenced at less than 1x coverage. Availability and implementationAlphaFamImpute is available as a Python package from the AlphaGenes website, http://www.AlphaGenes.roslin.ed.ac.uk/AlphaFamImpute. Contactawhalen@roslin.ed.ac.uk Supplementary informationA complete description of the methods is available in the supplementary information.

genetics↗

bioRxiv2019

Origin Specific Genomic Selection: a simple process to optimize the favourable contribution of parents to progeny

Modern crop breeding is in constant demand for new genetic diversity as part of the arms race with genetic gain. The elite gene pool has limited genetic variation and breeders are trying to introduce novelty from unadapted germplasm, landraces and wild relatives. For polygenic traits, currently available approaches to introgression are not ideal, as there is a demonstrable bias against exotic alleles during selection. Here, we propose a partitioned form of genomic selection, called Origin Specific Genomic Selection (OSGS), where we identify and target selection on favourable exotic alleles. Briefly, within a population derived from a bi-parental cross, we isolate alleles originating from the elite and exotic parents, which then allows us to separate out the predicted marker effects based on the allele origins. We validated the usefulness of OSGS using two nested association mapping (NAM) datasets: barley NAM (elite-exotic) and maize NAM (elite-elite), as well as by computer simulation. Our results suggest that OSGS works well in bi-parental crosses, and it is possible to extend the approach to broader multi-parental populations.

genetics↗

bioRxiv2019

Evaluation of sequencing strategies for whole-genome imputation with hybrid peeling

BackgroundFor assembling large whole-genome sequence datasets to be used routinely in research and breeding, the sequencing strategy should be adapted to the methods that will later be used for variant discovery and imputation. In this study we used simulation to explore the impact that the sequencing strategy and level of sequencing investment have on the overall accuracy of imputation using hybrid peeling, a pedigree-based imputation method well-suited for large livestock populations.\n\nMethodsWe simulated marker array and whole-genome sequence data for fifteen populations with simulated or real pedigrees that had different structures. In these populations we evaluated the effect on imputation accuracy of seven methods for selecting which individuals to sequence, the generation of the pedigree to which the sequenced individuals belonged, the use of variable or uniform coverage, and the trade-off between the number of sequenced individuals and their sequencing coverage. For each population we considered four levels of investment in sequencing that were proportional to the size of the population.\n\nResultsImputation accuracy largely depended on pedigree depth. The distribution of the sequenced individuals across the generations of the pedigree underlay the performance of the different methods used to select individuals to sequence. Additionally, it was critical to balance high imputation accuracy in early generations as well as in late generations. Imputation accuracy was highest with a uniform coverage across the sequenced individuals of around 2x rather than variable coverage. An investment equivalent to the cost of sequencing 2% of the population at 2x provided high imputation accuracy. The gain in imputation accuracy from additional investment diminished with larger populations and larger levels of investment. However, to achieve the same imputation accuracy, a proportionally greater investment must be used in the smaller populations compared to the larger ones.\n\nConclusionsSuitable sequencing strategies for subsequent imputation with hybrid peeling involve sequencing around 2% of the population at a uniform coverage around 2x, distributed preferably from the third generation of the pedigree onwards. Such sequencing strategies are beneficial for generating whole-genome sequence data in populations with deep pedigrees of closely related individuals.

genomics↗

bioRxiv2019

Accuracy of whole-genome sequence imputation using hybrid peeling in large pedigreed livestock populations

BackgroundWe demonstrate high accuracy of whole-genome sequence imputation in large livestock populations where only a small fraction of individuals (2%) had been sequenced, mostly at low coverage.\n\nMethodsWe used data from four pig populations of different sizes (18,349 to 107,815 individuals) that were broadly genotyped at densities between 15,000 and 75,000 markers genome-wide. Around 2% of the individuals in each population were sequenced (most at 1x or 2x and a small fraction at 30x; average coverage per individual: 4x). We imputed whole-genome sequence with hybrid peeling. We evaluated the imputation accuracy by removing the sequence data of a total of 284 individuals that had been sequenced at high coverage, using a leave-one-out design. We complemented these results with simulated data that mimicked the sequencing strategy used in the real populations to quantify the factors that affected the individual-wise and variant-wise imputation accuracies using regression trees.\n\nResultsImputation accuracy was high for the majority of individuals in all four populations (median individual-wise correlation was 0.97). Individuals in the earliest generations of each population had lower accuracy than the rest, likely due to the lack of marker array data for themselves and their ancestors. The main factors that determined the individual-wise imputation accuracy were the genotyping status of the individual, the availability of marker array data for immediate ancestors, and the degree of connectedness of an individual to the rest of the population, but sequencing coverage had no effect. The main factors that determined variant-wise imputation accuracy were the minor allele frequency and the number of individuals with sequencing coverage at each variant site. These results were validated with the empirical observations.\n\nConclusionsThe coupling of an appropriate sequencing strategy and imputation method, such as described and validated here, is a powerful strategy for generating whole-genome sequence data in large pedigreed populations with high accuracy. This is a critical step for the successful implementation of whole-genome sequence data for genomic predictions and fine-mapping of causal variants.

genomics↗

bioRxiv2019

Efficient use of genomic information for sustainable genetic improvement in small cattle populations

This paper compares genetic gain, genetic variation, and the efficiency of converting variation into gain under different genomic selection scenarios with truncation or optimum contribution selection in a small dairy population by simulation. Breeding programs have to maximize genetic gain but also ensure sustainability by maintaining genetic variation. Numerous studies showed that genomic selection increases genetic gain. Although genomic selection is a well-established method, small populations still struggle with choosing the most sustainable strategy to adopt this type of selection. We developed a simulator of a dairy population and simulated a model after the Slovenian Brown Swiss population with ~10,500 cows. We compared different truncation selection scenarios by varying i) the method of sire selection and their use on cows or bull-dams, and ii) selection intensity and the number of years a sire is in use. Furthermore, we compared different optimum contribution selection scenarios with optimization of sire selection and their usage. We compared the scenarios in terms of genetic gain, selection accuracy, generation interval, genetic and genic variance, the rate of coancestry, effective population size, and the conversion efficiency. The results show that early use of genomically tested sires increased genetic gain compared to progeny testing as expected from changes in selection accuracy and generation interval. A faster turnover of sires from year to year and higher intensity increased the genetic gain even further but increased the loss of genetic variation per year. While maximizing intensity gave the lowest conversion efficiency, a faster turn-over of sires gave an intermediate conversion efficiency. The largest conversion efficiency was achieved with the simultaneous use of genomically and progeny tested sires that were used over several years. Compared to truncation selection optimizing sire selection and their usage increased the conversion efficiency by either achieving comparable genetic gain for a smaller loss of genetic variation or achieving higher genetic gain for a comparable loss of genetic variation. Our results will help breeding organizations to implement sustainable genomic selection.

genomics↗

bioRxiv2019

Modelling spatial variation in agricultural field trials with INLA

The objective of this paper was to fit different established spatial models for analysing agricultural field trials using the open-source R package INLA. Spatial variation is common in field trials and accounting for it increases the accuracy of estimated genetic effects. However, this is still hindered by the lack of available software implementations. Here we compare some established spatial models and show possibilities for flexible modelling with respect to field trial design and joint modelling over multiple years and locations. We use a Bayesian framework and for statistical inference the Integrated Nested Laplace Approximations (INLA) implemented in the R package INLA. The spatial models we use are the well-known independent row and column effects, separable first-order autoregressive (AR1{otimes}AR1) models and a geostatistical model using the stochastic partial differential equation (SPDE) approach. The SPDE approach models a Gaussian random field, which can accommodate flexible field trial designs and yields interpretable parameters. We test the models in a simulation study imitating a wheat breeding program with different levels of spatial variation, with and without genome-wide markers, and with combining data over two locations, modelling spatial and genetic effects jointly. We evaluate predictive performance by correlation between true and estimated breeding values, the continuous rank probability score and how often the best individuals rank at the top. The results show best predictive performance with the AR1{otimes}AR1 and the SPDE. We also present an example of fitting the models to real wheat breeding data and simulated tree breeding data with the Nelder wheel design.\n\nKey messageEstablished spatial models improve the analysis of agricultural field trials with or without genomic data and can be fitted with the open-source R package INLA.

genomics↗

bioRxiv2018

A family-based phasing algorithm for sequence data

This paper describes a family-based phasing algorithm, for variable-coverage sequence data, that first minimises phasing errors and then maximises the proportion of alleles phased. This algorithm is one of the essential tools that underpin an overall strategy for generating highly accurate sequence data on whole populations at low cost.\n\nThe algorithm is called AlphaFamSeq. It uses sequence data on the focal individual and at least two generations of ancestors to phase alleles. In the first step, AlphaFamSeq calculates allele probabilities using iterative peeling. In subsequent steps, the alleles are phased using heuristics deriving information from the sequence data of parents, grandparents and progenies and, if available, from other families in the pedigree. AlphaFamSeq was tested on a range of simulated data sets.\n\nAlphaFamSeq gives low phasing error rates and, if there is sufficient sequence information and haplotype sharing amongst individuals, it can give a high yield of correctly phased alleles.\n\nThe allele threshold had a large effect and window size had a small effect on performance. When all individuals in a single family were sequenced at different coverages the highest correctly phased alleles reached 90% of the possible maximum (98.9%) at ~1/6 of the maximum aggregate coverage. Adding sequence information from other related individuals increased the percentage of correctly phased alleles. Imputation performance was high across all allele frequencies (average correlation by marker of 0.94), except for a slight decrease at very low frequencies ([≤]0.01 MAF).\n\nWithin an overall strategy for generating highly accurate sequence data on whole populations at low cost the role of AlphaFamSeq is to provide very accurately phased haplotypes on focal individuals, who are individuals whose haplotypes are very common in the population.

genomics↗

bioRxiv2018

Plant breeders should be determining economic weights for a selection index instead of using independent culling for choosing parents in breeding programs with genomic selection

In the context of genomic selection, we evaluated and compared recurrent selection breeding programs using either index selection or independent culling for selection of parents. We simulated a clonally propagated crop breeding program for 20 cycles of selection using either independent culling or an economic selection index with two unfavourably correlated traits under selection. Cycle time from crossing to selection of parents was kept the same for both strategies. Our results demonstrate that accurate knowledge of the economic importance of traits is essential even when performing independent culling. This is because independent culling achieved its optimum genetic gain when the culling threshold for each trait varied accordingly to the economic importance of the traits. When gains from independent culling were maximised, the efficiency of converting genetic diversity into genetic gain of both selection methods were equivalent. When the same proportion selected of 10% for each trait was used instead of optimal culling levels, index selection was 10%, 128% and 310% more efficient than independent culling when T2 had a relative economic importance of 1.0, 2.5 and 5.0, respectively. Given the complexity of estimating optimal culling levels and the fact that the gains achieved with independent culling are, at most, equivalent to index selection, the use of an economic selection index is recommended for multi-trait genomic selection.

genetics↗

bioRxiv2018

Family-specific genotype arrays increase the accuracy of pedigree based imputation at very low marker densities

In this paper we evaluate the performance of using a family-specific low-density genotype arrays to increase the accuracy of pedigree based imputation. Genotype imputation is a widely used tool that decreases the costs of genotyping a population by genotyping the majority of individuals using a low-density array and using statistical regularities between the low-density and high-density individuals to fill in the missing genotypes. Previous work on population based imputation has found that it is possible to increase the accuracy of imputation by maximizing the number of informative markers on an array. In the context of pedigree based imputation, where the informativeness of a marker depends only on the genotypes of an individuals parents, it may be beneficial to select the markers on each low-density array on a family-by-family basis. In this paper we examined four family-specific low-density marker selection strategies, and evaluated their performance in the context of a real pig breeding dataset. We found that family-specific or sire-specific arrays could increase imputation accuracy by 0.11 at 1 marker per chromosome, by 0.027 at 25 markers per chromosome and by 0.007 at 100 markers per chromosome. These results suggest that there may be a room to use family-specific genotyping for very-low-density arrays particularly if a given sire or sire-dam pairing have a large number of offspring.

genetics↗

bioRxiv2018

Extending long-range phasing and haplotype library imputation algorithms to very large and heterogeneous datasets

BackgroundThis paper describes the latest improvements to the long-range phasing and haplotype library imputation algorithms that enable them to successfully phase both datasets with one million individuals and datasets genotyped using different sets of single nucleotide polymorphisms (SNPs). Previous publicly available implementations of long-range phasing could not phase large datasets due to the computational cost of defining surrogate parents by exhaustive all-against-all searches. Further, both long-range phasing and haplotype library imputation were not designed to deal with large amounts of missing data, which is inherent when using multiple SNP arrays.\n\nMethodsHere, we developed methods which avoid the need for all-against-all searches by performing long-range phasing on subsets of individuals and then combing results. We also extended long-range phasing and haplotype library imputation algorithms to enable them to use different sets of markers, including missing values, when determining surrogate parents and identifying haplotypes. We implemented and tested these extensions in an updated version of our phasing software AlphaPhase.\n\nResultsA simulated dataset with one million individuals genotyped with the same set of 6,711 SNP for a single chromosome took two days to phase. A larger dataset with one million individuals genotyped with 49,579 SNP for a single chromosome took 14 days to phase. The percentage of correctly phased alleles at heterozygous loci was respectively 90.5% and 90.0% for the two datasets, which is comparable to the accuracy achieved with previous versions of AlphaPhase on smaller datasets.\n\nThe phasing accuracy for datasets with different sets of markers was generally lower than that for datasets with one set of markers. For a simulated dataset with three sets of markers 2.8% of alleles at heterozygous positions were phased incorrectly whereas the equivalent figure with one set of markers was 0.6%.\n\nConclusionsThe improved long-range phasing and haplotype library imputation algorithms enable AlphaPhase to quickly and accurately phase very large and heterogeneous datasets. This will enable more powerful breeding and genetics research and application.

bioinformatics↗

bioRxiv2018

The effects of training population design on genomic prediction accuracy in wheat

Genomic selection offers several routes for increasing genetic gain or efficiency of plant breeding programs. In various species of livestock there is empirical evidence of increased rates of genetic gain from the use of genomic selection to target different aspects of the breeders equation. Accurate predictions of genomic breeding value are central to this and the design of training sets is in turn central to achieving sufficient levels of accuracy. In summary, small numbers of close relatives and very large numbers of distant relatives are expected to enable accurate predictions.\n\nTo quantify the effect of some of the properties of training sets on the accuracy of genomic selection in crops we performed an extensive field-based winter wheat trial. In summary, this trial involved the construction of 44 F2:4 bi- and triparental populations, from which 2992 lines were grown on four field locations and yield was measured. For each line, genotype data were generated for 25,000 segregating single nucleotide polymorphism markers. The overall heritability of yield was estimated to 0.65, and estimates within individual families ranged between 0.10 and 0.85. Within cross genomic prediction accuracies of yield BLUEs were 0.125 - 0.127 using two different cross-validation approaches, and generally increased with training set size. Using related crosses in training and validation sets generally resulted in higher prediction accuracies than using unrelated crosses. The results of this study emphasize the importance of the training set design in relation to the genetic material to which the resulting prediction model is to be applied.

genetics↗