Genomic prediction of crown rust resistance in Lolium perenne

Genomic prediction of crown rust resistance in Lolium perenne Background: Genomic selection (GS) can accelerate genetic gains in breeding programmes by reducing the time it takes to complete a cycle of selection. Puccinia coronata f. sp lolli (crown rust) is one of the most widespread diseases of perennial ryegrass and can lead to reductions in yield, persistency and nutritional value. Here, we used a large perennial ryegrass population to assess the accuracy of using genome wide markers to predict crown rust resistance and to investigate the factors affecting predictive ability. Results: Using these data, predictive ability for crown rust resistance in the complete population reached a maximum of 0.52. Much of the predictive ability resulted from the ability of markers to capture genetic relationships among families within the training set, and reducing the marker density had little impact on predictive ability. Using permutation based variable importance measure and genome wide association studies (GWAS) to identify and rank markers enabled the identification of a small subset of SNPs that could achieve predictive abilities close to those achieved using the complete marker set. Conclusion: Using a GWAS to identify and rank markers enabled a small panel of markers to be identified that could achieve higher predictive ability than the same number of randomly selected markers, and predictive abilities close to those achieved with the entire marker set. This was particularly evident in a sub-population characterised by having on-average higher genome-wide linkage disequilibirum (LD). Higher predictive abilities with selected markers over random markers suggests they are in LD with QTL. Accuracy due to genetic relationships will decay rapidly over generations whereas accuracy due to LD will persist, which is advantageous for practical breeding applications. Keywords: Genomic selection, Crown rust, Perennial ryegrass, Genetic relationship, GWAS Background that can be utilized to develop resistant cultivars [11–13]. Perennial ryegrass (Lolium perenne L.) is the predominant Phenotypic recurrent selection is typically used to develop forage species grown in temperate regions of the world [1]. cultivars with improved resistance and selection is often Puccinia coronata f. sp. lolli (crown rust) is one of the most carried out on spaced plants [9, 11, 12, 14]. There is a high widespread diseases of perennial ryegrass and can lead to correlation between spaced plants and swards for disease a reduction in forage nutritive value, yield and persistency resistance and indirect selection for disease resistance [2–4]. Poor quality, rust infected swards can impact ani- on spaced plants can improve resistance in sward condi- mal performance and well-being [5–7]. Developing resis- tions [15]. However, with the advancements in molecu- tant cultivars is the most viable option for disease control lar marker development over the last decade, efforts to and it has been shown that resistance to crown rust is use marker assisted breeding strategies have been pur- conferred by both quantitative and qualitative inheritance sued. One such strategy involves identifying quantitative [8–11]. As an obligate out-crossing species, perennial rye- trait loci (QTL) in bi-parental mapping populations and grass germplasm has high variation for disease resistance using markers to efficiently backcross the QTL into elite breeding material [16]. Although QTLs explaining signif- *Correspondence: stephen.byrne@teagasc.ie icant phenotypic variation for crown rust resistance were Teagasc, Crop Science Department, Oak Park, R93 XE12 Carlow, Ireland mapped onto linkage group (LG) 1-5 and 7 [17–23], it is Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Arojju et al. BMC Genetics (2018) 19:35 Page 2 of 10 unclear if any of these QTLs were successfully introduced a glasshouse and later transplanted to the field in 2013 at into breeding material. Genome wide association studies Oak Park, Carlow, Ireland (52° 51 34.2 N,6°55 03.0 W). (GWAS) are another approach to identify markers linked Plants were grown in two replicates in a partially balanced to QTL. In this case breeding populations can directly be incomplete block design. Each block consists of 60 test used to identify marker-trait associations, although iden- genotypes and 5 check genotypes and was surrounded tified markers tended to explain a small proportion of the by a 1.5 m sward consisting of a four way mix of crown total additive genetic variance, resulting in smaller genetic rust susceptible perennial ryegrass cultivars. Crown rust gains [24–26]. was recorded in the years 2014 and 2015 as mean per- Genomic selection (GS) was first proposed by Meuwissen centage disease score on each plant. Briefly, percentage et al. [27], as a method to capture complete additive disease score was obtained by combining scores of per- genetic variance using genome wide markers. GS is a form centage of leaves with infection and average percentage of of marker assisted breeding, which accounts for all marker infection on diseased leaves. Scoring was done at multi- effects across the entire genome to calculate genomic esti- ple time points in September to November without any mated breeding values (GEBVs), which are used to select harvest cuts between scorings (Table 1). We are trying individual plants for advancement [26]. Use of genome- to develop genomic models to identify plants with good wide markers will include small effect loci and is ideal resistance to crown rust across the season, and we decided for complex traits with low to moderate heritability. In to use all time points for constructing a quantitative sum- GS, a training population is genotyped with genome wide mary for crown rust resistance. To do this we calculated markers and phenotyped for the trait under selection AUDPC for each spaced plant in both years. Using mul- and models to predict breeding values from marker data tiple time point data, we generated area under disease are developed. Implementing GS for complex traits like progress curve (AUDPC) as follows: yield and quality is a primary objective of many peren- N −1 (y + y ) i i+1 nial ryegrass breeding programmes. In contrast to yield A = (t − t ) (1) k i+1 i and quality traits, the cost (labour and time) of pheno- i=1 typing for disease resistance is much lower. However, it where y is the extent of infection (percentage disease is important that any GS approaches targeting yield and th score) at i observation and t isthetimepoint at quality improvements also ensure adequate disease resis- th i observation. N is the number of individuals in the tance is maintained, particularly where multiple rounds data set. of marker based selections are performed between field Variance components for crown rust were estimated evaluations. Opportunities for GS in perennial ryegrass using the R package lme4 (linear mixed-effects models were first reviewed by Hayes et al. [28], and the earli- using ’eigen’ and S4) [34]. Broad sense heritability was est empirical study was done by Fè et al. [29] for heading estimated as follows: date, which confirmed the superiority of GS over marker assisted selection. Later Fè et al. [30], Grinberg et al. [31] g H = (2) and Byrne et al. [32] reported high predictive ability for 2 2 2 σ + σ /2 + σ /4 g g∗yr res important agronomical traits in perennial ryegrass. In particular, predictive ability for crown rust reached up to where σ is the total genetic variance among individu- 0.58 [30] when genotypes and phenotypes were evaluated als, σ is the variance associated with genotype by year g∗yr on F families. In this study, we evaluated predictive abil- ity for crown rust resistance on individual plants in a large Table 1 Mean percentage disease score for crown rust resistance perennial ryegrass population, and assessed factors con- at different time points (TP) in Year1 (2014) and Year2 (2015) tributing to predictive ability, such as training population Time point/dates Mean SD Min Max size and marker density. We also performed a GWAS to identify a small to moderately sized panel of markers with Year 1 good predictive ability for crown rust resistance. TP1 (13/10/14) 3.1 6.1 0 40 TP2 (20/10/14) 5.2 7.6 0 45 Methods TP3 (29/10/14) 9.6 10.8 0 60 Plant material, phenotyping and genotyping TP4 (10/11/14) 9.8 8.7 0 45 The training population consists of 30 diploid peren- nial ryegrass families that have been described previously Year 2 [32, 33]. Each family consists of 60 genotypes making up a TP1 (21/09/15) 2.0 4.4 0 32 population of 1800 individuals. The complete population TP2 (05/10/15) 11.2 10.0 0 60 consists of ten cultivars, eight full-sib families, eight half- TP3 (19/10/15) 19.9 9.0 0 63 sib families and four ecotypes. Plants were established in Arojju et al. BMC Genetics (2018) 19:35 Page 3 of 10 interaction and σ is residual variance. With genotype as repeated 100 times and the resulting accuracies were then res random effect and year and checks as fixed effects, condi- averaged. This approach to cross-validation has previ- tional modes (BLUPs) were calculated in lme4 and used as ously been used to evaluate genomic prediction mod- input for genomic prediction. els [42, 43]. Predictive ability and bias were assessed Genotyping was done using genotyping by sequenc- in the complete population and in each sub-population. ing (GBS) approach described by Elshire et al. [35]and Predictive ability (r ) was determined as the Pearson’s data analysed as described in Byrne et al. [32]. Briefly, correlation coefficient between observed phenotypic genomic DNA was extracted from leaf samples and GBS value and predicted phenotype. Bias was evaluated by libraries were prepared using the restriction enzyme regressing observed phenotypic value on predictions. ApeKI, libraries were amplified and sequenced on an Illu- We reduced training population size and marker den- mina Hiseq2000. Panels of SNPs were identified in the sity in order to identify the impact of training population complete population, as well as in all sub populations size and marker number on predictive ability. To com- separately (half-sibs, full-sibs, ecotypes, cultivars). Indi- pare predictive ability for traits with contrasting genetic viduals with very low sequencing coverage and/or largely architecture we compared heading date, a highly heri- missing phenotypic data were eliminated from the anal- table trait, with crown rust. Predictive ability for head- ysis giving a final population for analysis of 1582 indi- ing date has already been shown to be high (0.81) in viduals. Missing marker data was imputed using mean this population [32]. We re-analyzed data for heading imputation. date according to methods described above and made a comparison with crown rust. To evaluate the impact of Genomic prediction models leaving related material out of the training set we also We used four statistical algorithms for genomic pre- performed cross validation by leaving one family out. In diction, ridge regression best linear unbiased prediction this approach one complete family (up to 60 individuals) (rrBLUP) [27], Bayes B [36] and Bayesian Lasso [37], and is left out of the training set and only used for testing. random forest [38]. rrBLUP is a mixed model approach, This was repeated so that each family in turn is used as a which was initially proposed for GS. We used an R pack- test set. age called rrBLUP [39] for fitting the mixed model as follows Genome wide association A mixed linear model (MLM) was also used for associ- y = μ + Xg +  (3) ation mapping, implemented in the R package rrBLUP where μ istheoverall mean,Xisthemarkermatrix, gis [39]. Population structure and family relatedness was accounted for with a kinship matrix calculated by rrBLUP the matrix of marker effects,  is a vector of residual from the input genotypic data. We accounted for multiple effects and y is a vector of conditional modes for crown testing using a Bonferroni correction and markers pass- rust. We also evaluated two Bayesian approaches, Bayes B ing an α level 0.05 threshold were considered statistically [36] and Bayesian Lasso [37], which were both imple- significant. mented using the R package BGLR [40] with the following parameters: number of iterations = 5000, burn-in = 500 and thinning = 5. Random forest is a machine-learning Results and discussion tool, in which series of regression trees were grown inde- Phenotypic analysis for crown rust pendently to the largest extent possible using subsets of The mean percentage disease score for crown rust infec- bootstrap samples. At each split of the tree, a random tion in the population increased over time in both eval- subset of variables is selected to identify the best split. uation years as infection levels accumulated (Table 1). In We implemented random forest using the R package ran- both years, evaluations were carried out in the period domForest [41], setting the number of variables at each from September to November during a time when disease split to 1/3 of the total variables, and using a terminal pressure tends to be at its greatest [15, 44]. The highest node size of five and minimum of 500 trees per forest. We mean percentage disease score was seen in late October also used random forest to rank variables using the vari- 2015 and was more than double the highest mean percent- able importance measure, a permutation based measure age disease score from 2014 (Table 1). In addition to plant in which variables are ranked based on the mean decrease health and level of host resistance, crown rust infection is in accuracy. influenced by various environmental factors, such as tem- perature, relative humidity, and light [45–47]. The latency Cross validation scheme period is reduced and spore production increased as tem- We evaluated genomic prediction models using Monte- perature increases [45], and it has been shown that when Carlo cross-validation by randomly assigning plants into temperatures exceed 25°C, the susceptibility of previously training (70%) and test (30%) sets and the procedure was resistant cultivars can be increased [46]. It has already Arojju et al. BMC Genetics (2018) 19:35 Page 4 of 10 been shown that there is variability within pathogen pop- and used to aid selection of the top performing families ulations, and different races can be found within and from which to construct the synthetics cultivars. During between locations. It is also possible that the composition construction of synthetics a spaced plant nursery may be of a pathogen population can change over short periods of established to evaluate heading date and crown rust resis- time and plants that are resistant at one point in time will tance before selecting individual genotypes from which to become susceptible as the pathogen population shifts or construct synthetics (within family selection). In practice, evolves. this has a time cost of 2 to 3 years (establishment, eval- AUDPC values ranged from 0 to 1371 and the Pearson uation, selection and recombining), and using molecular correlation co-efficient between replicates within years markers offers an opportunity to reduce this to one year was moderate (0.69 in 2014 and 0.59 in 2015). However, in those selection cycles where GEBVs are predicted. This the Pearson correlation co-efficient between years was depends on our ability to accurately predict traits such as low (0.28), and there was a significant genotype by year crown rust from genomic data. interaction (F = 3.025, MSE = 60676, p = 0.0001). (1761) The broad sense heritability for crown rust infection Predicting crown rust resistance with genomic data was moderate (0.36), which is in line with previous We evaluated four algorithms for prediction of crown rust estimates of heritability calculated in other populations infection from genomic data, rrBLUP, Bayes B, Bayesian [11, 48]. Overall there is a good phenotypic variation Lasso, and random forest. The mean predictive ability for crown rust infection among and within the 30 fam- after cross-validation within the complete population was ilies/cultivars/ecotypes making up the entire population 0.52 using rrBLUP, 0.52 using Bayesian Lasso, 0.51 using (Fig. 1). Plants were placed into one of four categories Bayes B, and 0.49 using random forest (Additional file 1: (sub-populations) based on mating type or origin, these Figure S1). rrBLUP was computationally faster, and there- were (i) full-sib families, (ii) half-sib families, (iii) cultivars, fore results from all further analysis are only reported and (iv) ecotypes. In general the ecotypes were more sus- for rrBLUP. The predictive ability of 0.52 is in line with ceptible to crown rust infection than cultivars or breed- previous estimates reported in perennial ryegrass where ing material (Fig. 1), which presents a challenge for the predictions were based on mean genotypes and pheno- incorporation of ecotypes into breeding programmes. The types of F families [30]. Predictive ability did not dif- broad-sense heritability calculated in each sub-population fer depending on whether the equations were developed varied between 0.17 in the cultivars to 0.44 in the full-sib using phenotypes from the last time point scored or the families. AUDPC values incorporating all time points. This indi- Crown rust infection is typically evaluated in breeding cates that a single scoring each year would have sufficed. programmes by growing spaced plants or potted plants However,theimportanceofevaluatingcrown rust in more from a population and visually scoring the level of crown than one year was emphasised by the low correlation rust infection. A mean score is assigned to each family between scores in 2014 and 2015. When we calculate the predictive ability within each of the sub-populations (cultivars, half-sib families, full-sib families, and ecotypes), the highest predictive ability for crown rust was obtained using plants from full-sib fami- lies (0.54) and the lowest predictive ability for crown rust was obtained with the plants from the ecotypes (0.24) (Fig. 2). Generally, traits with higher heritability achieve higher predictive abilities [49, 50], and we see that here where crown rust measurements taken in the full-sib families had the highest broad-sense heritability and the highest predictive ability. In general, there was a good cor- relation between predictive ability and both phenotypic variance and heritability. This relationship between phe- notypic variance and predictive ability has been observed previously [49, 51]. We also evaluated the predictive ability using a leave- one-family-out cross validation scheme. The complete Fig. 1 Phenotypic variation for crown rust resistance in the complete population, grouped according to sub-population types: cultivars population is comprised of 30 families/cultivars/ecotypes, (CS), ecotypes (ES), full-sibs (FS) and half-sibs (HS). Broad sense each with up to 60 individual genotypes. The predic- heritability (H ) in complete population and sub-populations is tive ability was assessed in the complete population by highlighted over the figure selectively leaving one family out of the training set and Arojju et al. BMC Genetics (2018) 19:35 Page 5 of 10 Fig. 2 Predictive ability in different population types. Complete population (CP), cultivars (CS), ecotypes (ES), full-sibs (FS) and half-sibs (HS) are listed on x-axis, predictive ability (left) and bias (right) on y-axis. Crown rust is in red and heading date in blue using it for testing. In addition to crown rust we also The drop in predictive ability was more pronounced as we evaluated predictive ability for heading date phenotypes reduced the training population size for crown rust resis- previously reported [32]. The predictive ability for both tance than it was for heading date. The predictive ability crown rust (r = 0.02, min =−0.36, max = 0.36) and for crown rust resistance when using 90% of the popula- heading date (r = 0.29, min =−0.14, max = 0.65) varied tion as a training set was 0.52 and the predictive ability was greatly depending on which family was left out, and hav- 0.38 when using just 10% of the population. Irrespective ing related material in the training set (shared parentage) of the trait, as the training population size increased there greatly improved predictive ability. was an increase in predictive ability which is consistent with similar correlations between training population size Effect of training population size and marker density on and predictive ability reported previously for perennial predictive ability ryegrass [29, 30]and othercrops [51–54]. Useful linkage As we reduced the number of individuals in the train- disequilibrium (LD) only extends over short distances in ing population we saw a decrease in predictive ability and perennial ryegrass and it has been suggested that this is an increasingly upward bias in the variance of predictions the result of a very large past effective population size [28]. This impacts both the size of the reference population for both crown rust resistance and heading date (Fig. 3). Fig. 3 Effect of training population size on predictive ability. Training population is varied from 90% (1423 individuals) to 10% (158 individuals) on x-axis and predictive ability (left), bias (right) on y-axis. Crown rust is in red and heading date in blue Arojju et al. BMC Genetics (2018) 19:35 Page 6 of 10 and marker density required to achieve high accuracies Identifying SNPs associated with crown rust resistance when predicting traits from genomic data. The fact that The cost of genotyping impacts the number of selection we are able to achieve high predictive abilities with rela- candidates that can be evaluated and therefore impacts tively small training populations is likely a result of strong the selection intensity. Different approaches to low den- genetic structure and differentiation in our diverse popu- sity SNP genotyping for genomic selection have been lation and the use of the marker data to capture genetic proposed. These include variable selection methods to relationships [55]. identify a small subset of markers in strong LD with the The limited LD also affects the number of markers trait [56] or using a small random subset of markers to required to obtain high predictive accuracies, and given impute from low-to-high density [57]. Until a chromo- the extent of LD in the broader perennial ryegrass pop- some scale assembly of the perennial ryegrass genome ulation, marker numbers in excess of one million have becomes available the latter remains a challenge. We used been suggested for achieving high accuracies [28]. When both permutation based variable importance measures we reduced marker number in the complete population and GWAS analysis to identify a subset of markers capa- and the various sub-populations we observed very little ble of predicting crown rust resistance. Using permutation impact on the predictive ability for either trait (Table 2). basedvariableimportancemeasureswewereabletorank Reducing the marker set to 5% of the total available had markers by mean decrease in accuracy and select the top virtually no impact on predictive ability in all cases. This ranked markers for use in genomic prediction. In the would support our observation that much of the predic- case of GWAS we ranked SNPs based on significance and tive ability can be derived from makers capturing familial again selected the top ranked markers for use in genomic relationships. When marker number dropped below 5% prediction. All variable importance measures and GWAS (10878) predictive ability for both traits in the complete were identified and ranked in the training set and used population began to drop. However, even with 0.05% to predict phenotypes in the test set via cross-validation. (109) of markers the mean predictive ability was 0.30 for When we used the top 100 ranked markers from the per- crown rust resistance and 0.52 for heading date. Know- mutation based variable importance measures, the mean ing the contribution of genetic relationships to predictive predictive ability of 100 iterations was 0.42 (ranging from ability is important because it will change over genera- 0.36 to 0.48). When we used the top 100 ranked markers tions. In contrast, predictive ability due to LD has greater from the GWAS analysis, the mean predictive ability of persistence over generations and is therefore preferen- 100 iterations was 0.36 (ranging from 0.25 to 0.44). In both tial [55]. Schemes for implementing genomic selection cases the mean predictive ability with selected markers in perennial ryegrass that pursue a reduction in effec- is higher than the predictive ability with random mark- tive population size from the outset have been proposed. ers, which was 0.28 (ranging from 0.18 to 0.39). The lower Such schemes would lead to an increase in the extent predictive ability using GWAS marker selection is not of LD and ensure that predictive ability due to LD can surprising considering that we corrected for population be captured using a reasonable number of markers and structure using a kinship matrix, and we are more reliant a reference population size that is feasible in breeding on identifying markers in LD with the trait. As discussed programmes. above, the predictive ability of these markers is expected Table 2 Predictive ability (r ) and bias for crown rust (CR) and heading date (HD) by selecting random markers of 100 to 0.05%, in complete population (CP), cultivars (CS), full-sibs (FS) and half-sibs (HS) Pop 100% 60% 20% 5% 1% 0.5% 0.1% 0.05% r bias r bias r bias r bias r bias r bias r bias r bias p p p p p p p p CR CP 0.52 1.22 0.52 1.22 0.52 1.21 0.51 1.18 0.46 1.10 0.43 1.07 0.36 1.04 0.30 1.48 CS 0.29 1.28 0.28 1.26 0.28 1.24 0.27 1.18 0.22 0.97 0.17 0.80 0.14 0.95 0.10 1.13 FS 0.54 1.13 0.54 1.13 0.54 1.13 0.54 1.14 0.52 1.07 0.50 1.03 0.45 1.00 0.40 0.99 HS 0.49 1.24 0.49 1.24 0.49 1.24 0.49 1.24 0.48 1.23 0.46 1.21 0.42 1.23 0.36 1.22 HD CP 0.81 1.16 0.81 1.16 0.81 1.16 0.80 1.14 0.75 1.07 0.72 1.05 0.62 1.01 0.52 1.00 CS 0.84 1.25 0.81 1.19 0.81 1.20 0.81 1.18 0.78 1.11 0.77 1.12 0.66 1.03 0.56 1.02 FS 0.76 1.00 0.75 1.16 0.75 1.16 0.75 1.16 0.74 1.14 0.68 1.27 0.64 1.26 0.62 0.87 HS 0.74 1.18 0.74 1.09 0.74 1.10 0.74 1.09 0.73 1.08 0.72 1.15 0.67 1.10 0.62 1.09 Arojju et al. BMC Genetics (2018) 19:35 Page 7 of 10 to be more persistent over subsequent generations. Using All markers were located within 16 genomic scaffolds GWAS selectedmarkersitisclear to seethattheyare containing 44 predicted genes. Out of 16 scaffolds we superior to randomly selected markers up to the point, were able to place seven scaffolds onto LG3, 5 and 7 beyond which adding more markers does not improve (Additional file 3: Table S2). We found five common scaf- predictive ability in either case (Fig. 4). The ability of a folds between the complete population and the IBERS GWAS within each sub-population to identify and select and only two of these scaffolds were mapped, onto LG3. a small set of SNPs with excellent predictive ability var- On LG3 five markers were anchored within 60.4-61.21 ied, and in some cases was little better than random SNP cM. Genes present on these scaffolds were coding for selection (Fig. 5). The GWAS on plants originating from domains including Mon1, Aquaporin, DUF1635, Nucleo- IBERs bred cultivars identified a small set of twenty SNPs redoxin, Beta-glucan export ATP-binding/permease pro- with 77% of the predictive ability achieved with 20,000 tein, BRASSINOSTEROID INSENSITIVE 1-associated SNPs. The power of a GWAS to identify markers with receptor kinase 1, Alpha N-terminal protein methyltrans- high predictive ability was much greater within the popu- ferase 1. Gene function of these domains plays a key role in lation made up of IBERs plants than within cultivars, and ATP-binding, membrane proteins, enzyme catalysis and full-sib families where twenty SNPs could only achieve 46 pathogen-associated molecular pattern (PAMP)-triggered and 48% of the predictive ability with 20,000 SNPs, respec- immunity (PTI) (Additional file 4:TableS3)[60]. tively. On average LD is higher within the sub-population Using small subsets of trait associated markers may be with IBERs plants, which may explain the greater an effective strategy for within-family prediction of traits ability to identify markers associated with crown-rust such as heading date, crown rust resistance and some resistance. quality traits. Predicting heading date from markers would In order to characterise the markers associated with enable plants to be matched in heading date to ensure crown rust resistance we repeated the GWAS analysis sufficient cross-pollination when constructing synthetic without division of genotypes into training and testing cultivars [32]. Combining these with markers to predict sets. We carried out GWAS using the complete pop- crown rust resistance would also avoid the inclusion of ulation and found 29 markers significantly associated plants with high levels of susceptibility, and furthermore with crown rust resistance after correction for multiple prediction models can be based on multi-year evaluations. testing (Additional file 2: Table S1). Using the peren- It is clear from the phenotypic data presented here that nial ryegrass genome [58] as a reference, we located all there is substantial within family variation for crown rust markers within 22 genomic scaffolds that contained 50 resistance. Opportunities already exist to genotype small predicted genes. Using the Genome Zippper [58, 59], to moderate sized marker panels in 1000s of samples at we anchored ten scaffolds onto LG2, 3, 4, 5 and 7 low cost [61]. Using these approaches small fragments (Additional file 3: Table S2). Similarly, we did GWAS (200-300 bp) are amplified and sequenced at hundreds of on IBERS material and found 24 markers associated loci. These amplicons can be used as short haplotypes in with crown rust resistance (Additional file 2:TableS1). marker aided selection strategies. An assay can be devel- oped to target loci in linkage with QTL for heading date [32], crown rust resistance, and other traits such as qual- ity parameters. The assay can also include a suite of loci with a good distribution throughout the genome and be deployed for among-and-within-full-sib-family selection (Additional file 5: Figure S2). Once high yielding fami- lies are identified in field trials, within family selection for crown rust resistance and forage quality can be per- formed at a high selection intensity with the molecular marker assay. Furthermore, plants can be selected to be synchronous in flowering time. Conclusions Our findings show that predicting crown rust resistance in perennial ryegrass can be achieved with high accuracy using AUDPC scores on spaced plants. However, there was no difference in predictive ability when equations Fig. 4 Predictive ability of selected markers versus random markers in the complete population. Markers were selected based on the were developed using phenotypes from the last time-point ranking from genome wide association studies and compared with scored or the AUDPC values, meaning a single time point random markers of similar size was adequate to evaluate the crown rust susceptibility of Arojju et al. BMC Genetics (2018) 19:35 Page 8 of 10 Fig. 5 Comparing predictive ability of selected versus random markers. Markers were selected based on the ranking from genome wide association studies in cultivars, full-sibs and IBERS material and compared with random markers of similar size the spaced plants. Much of the predictive ability comes Abbreviations AUDPC: Area under disease progress curve; GEBV: Genomic estimated from markers capturing familial relationships, highlighted breeding value; GS: Genomic selection; GWAS: Genome wide association by the observation that there was no drop in predic- study; LD: Linkage disequilibrium; QTL: Quantitative trait locus tive ability when going from the entire marker set down Acknowledgments to only 5% (10,878) of the marker set. Accuracy due Authors would like to thank Jean-Baptiste Enjelvin, Olivia Aylesbury, Mary to genetic relationships will decay rapidly over genera- O’Sullivan, Padraig Dunne and Michael Murphy for help with phenotyping and Sean Murray and Helena Meally for genomic DNA extraction. tions whereas accuracy due to LD will persist. Using a GWAS we attempted to identify and rank markers in Funding This work received funding from the Irish Department of Agriculture Food and LD with QTL. This enabled a small panel of markers to the Marine DAFM (RSF 11/S/109) and Teagasc core funding. SKA is supported be identified that had higher predictive ability than the by a Teagasc PhD Walsh Fellowship. SLB has received funding from the same number of randomly selected markers, and had pre- European Union’s Horizon 2020 research and innovation programme under dictive abilities close to those achieved with the entire the Marie Sklodowska-Curie grant agreement No. 658031. The funding bodies played no role in the design of the study and collection, analysis, and marker set. interpretation of data and in writing the manuscript. Availability of data and materials Additional files All sequence data associated with this project has been deposited in the NCBI Short Read Archive (SRA) under the BioProject accession PRJNA352789. Phenotypic data is available on Figshare (https://doi.org/10.6084/m9.figshare. Additional file 1: Figure S1. Predictive ability and bias for crown rust 5562640.v1), along with an R script template used for cross-validation (https:// using various algorithms for genomic prediction. (PDF 101 kb) doi.org/10.6084/m9.figshare.5712208.v1). Additional file 2: Table S1. List of markers associated with crown rust resistance based on genome wide association studies in complete Authors’ contributions population and IBERS material. (XLSX 14 kb) DM, PC, SB, TRH, MC conceived and designed the study. SKA and SLB performed the data analysis and drafted the initial manuscript. SKA, SLB, DM, Additional file 3: Table S2. List of genomic scaffolds where all the PC, TRH, MC, TM and SB contributed to interpretation of data and preparation significant markers from genome wide association studies were located. of the final manuscript. All authors read and approved the final version. Scaffolds were placed onto linkage group with the aid of Genome Zipper [59]. (XLSX 11 kb) Ethics approval and consent to participate Additional file 4: Table S3. List of predicted proteins on the genomic Not applicable. scaffolds. Markers located on these scaffolds were associated with crown rust resistance. BLAST was done on the predicted protein sequences using Competing interests PLAZA [62] to obtain the gene function. (XLSX 17 kb) The authors declare that they have no competing interests. Additional file 5: Figure S2. Among-and-within-full-sib-family selection Publisher’s Note that incorporates an inexpensive genotyping assay to implement Springer Nature remains neutral with regard to jurisdictional claims in within-family selection using a high selection intensity. (PDF 227 kb) published maps and institutional affiliations. Arojju et al. BMC Genetics (2018) 19:35 Page 9 of 10 Author details (Puccinia coronata f sp lolii) resistance in perennial ryegrass (Lolium Teagasc, Crop Science Department, Oak Park, R93 XE12 Carlow, Ireland. perenne) using AFLP markers and a bulked segregant approach. Teagasc, Grassland Science Research Department, Animal and Grassland Euphytica. 2005;143(1-2):135–44. Research and Innovation Centre, Oak Park, R93 XE12 Carlow, Ireland. 21. Muylle H, Baert J, Van Bockstaele E, Pertijs J, Roldàn-Ruiz I. Four QTLs Department of Agronomy, University of Wisconsin-Madison, WI53706 determine crown rust (Puccinia coronata f sp lolii) resistance in a perennial Madison, USA. Agricultural Research Service, United State Department of ryegrass (Lolium perenne) population. Heredity. 2005;95(5):348–57. Agriculture, WI53706 Madison, USA. Department of Botany, School of Natural 22. Schejbel B, Jensen LB, Xing Y, Lübberstedt T. QTL analysis of crown rust Sciences, Trinity College Dublin, Dublin 2 Ireland. resistance in perennial ryegrass under conditions of natural and artificial infection. Plant Breed. 2007;126(4):347–52. Received: 16 June 2017 Accepted: 18 April 2018 23. Thorogood D, Paget MF, Humphreys MO, Turner LB, Armstead I, Roderick H. QTL analysis of crown rust resistance in perennial ryegrass-implications for breeding. Int Turfgrass Soc Res J. 2001;9:218–23. 24. Hayes B, Goddard M. Genome-wide association and genomic selection References in animal breeding this article is one of a selection of papers from the 1. Humphreys M, Feuerstein U, Vandewalle M, Baert J. Ryegrasses. conference “exploiting genome-wide association in oilseed brassicas: a In: Fodder Crops and Amenity Grasses. New York: Springer; 2010. p. 211–60. model for genetic improvement of major OECD crops for sustainable 2. Potter LR. Effect of crown rust on regrowth, competitive ability and farming”. Genome. 2010;53(11):876–83. nutritional quality of perennial and Italian ryegrasses. Plant Pathol. 25. Heffner EL, Sorrells ME, Jannink J-L. Genomic selection for crop 1987;36(4):455–61. improvement. Crop Sci. 2009;49(1):1–12. 3. Valk P, Proveniers MCG, Pertijs JH, Lamers JTW, Dun CMP, Smeekens JCM. Late heading of perennial ryegrass caused by introducing an Arabidopsis 26. Jannink J-L, Lorenz AJ, Iwata H. Genomic selection in plant breeding: homeobox gene. Plant Breed. 2004;123(6):531–5. from theory to practice. Brief Funct Genom. 2010;9(2):166–77. 4. Price T. Ryegrass rust in victoria. Plant Protect Quart. 1987;2:189. 27. Meuwissen TH, Hayes BJ, Goddard ME. Prediction of total genetic value 5. Lancashire JA, Latch GCM. The effect of crown rust (Puccinia coronata using genome-wide dense marker maps. Genetics. 2001;157(4):1819–29. Corda) on the yield and botanical composition of two ryegrass/white 28. Hayes BJ, Cogan NO, Pembleton LW, Goddard ME, Wang J, clover pastures. N Z J Agric Res. 1970;13(2):279–86. Spangenberg GC, Forster JW. Prospects for genomic selection in forage 6. Plummer RM, Hall RL, Watt TA. The influence of crown rust (Puccinia plant species. Plant Breed. 2013;132(2):133–43. coronata) on tiller production and survival of perennial ryegrass (Lolium 29. Fè D, Cericola F, Byrne S, Lenk I, Ashraf BH, Pedersen MG, Roulund N, perenne) plants in simulated swards. Grass Forage Sci. 1990;45(1):9–16. Asp T, Janss L, Jensen CS, et al. Genomic dissection and prediction of 7. Smit HJ, Tas BM, Taweel HZ, Tamminga S, Elgersma A. Effects of heading date in perennial ryegrass. BMC Genomics. 2015;16(1):1. perennial ryegrass (Lolium perenne L,) cultivars on herbage production, 30. Fè D, Ashraf BH, Pedersen MG, Janss L, Byrne S, Roulund N, Lenk I, nutritional quality and herbage intake of grazing dairy cows. Grass Forage Didion T, Asp T, Jensen CS, et al. Accuracy of genomic prediction in a Sci. 2005;60(3):297–309. commercial perennial ryegrass breeding program. Plant Genome. 8. Hayward MD. Genetic control of resistance to crown rust (Puccinia 2016;9(3). coronato Corda) in Lolium perenne L, and its implications in breeding. 31. Grinberg NF, Lovatt A, Hegarty M, Lovatt A, Skøt KP, Kelly R, Blackmore T, Theor Appl Genet. 1977;51(2):49–53. Thorogood D, King RD, Armstead I, et al. Implementation of genomic 9. Kimbeng CA. Genetic basis of crown rust resistance in perennial ryegrass, prediction in Lolium perenne (L) breeding populations. Front Plant Sci. breeding strategies, and genetic variation among pathogen populations: 2016;7(133). a review. Anim Prod Sci. 1999;39(3):361–78. 32. Byrne S, Conaghan P, Barth S, Arojju SK, Casler M, Thibauld M, 10. McVeigh K. Breeding for resistance to crown rust (Puccinia coronata Velmurugan J, Milbourne D. Using variable importance measures to Corda var lolii Brown) in turf-type perennial ryegrass (Lolium perenne). identify a small set of SNPs to predict heading date in perennial ryegrass. New Brunswick: Rutgers University; 1975. PhD thesis. Sci Rep. 2017;7(3566). 11. Reheul D, Ghesquiere A. Breeding perennial ryegrass with better crown 33. Arojju SK, Barth S, Milbourne D, Conaghan P, Velmurugan J, Hodkinson TR, rust resistance. Plant Breed. 1996;115(6):465–9. Byrne SL. Markers associated with heading and aftermath heading in 12. Gibbs JG. Field resistance in Lolium sp to leaf rust (Puccinia coronata). perennial ryegrass full-sib families. BMC Plant Biol. 2016;16(1):160. Nature. 1966;209(5021):420. 34. Bates D, Mächler M, Bolker B, Walker S. Fitting linear mixed-effects 13. McGrath S, Hodkinson TR, Charles TM, Zen D, Barth S. Variation in models using lme4. arXiv preprint arXiv:14065823. 2014. inflorescence characters and inflorescence development in ecotypes and 35. Elshire RJ, Glaubitz JC, Sun Q, Poland JA, Kawamoto K, Buckler ES, cultivars of lolium perenne l. Grass Forage Sci. 2010;65(4):398–409. Mitchell SE. A robust, simple genotyping-by-sequencing (GBS) approach 14. Wilkins PW. Specialisation of crown rust on highly and moderately for high diversity species. PLoS ONE. 2011;6(5):19379. resistant plants of perennial ryegrass. Ann Appl Biol. 1978;88(1):179–84. 36. Habier D, Fernando RL, Kizilkaya K, Garrick DJ. Extension of the bayesian 15. Easton HS, Cooper BM, Frasers TJ, Widdup KH. Crown rust on perennial alphabet for genomic selection. BMC Bioinformatics. 2011;12(1):186. ryegrass in field trials. In: Proceedings of the New Zealand Grassland 37. Park T, Casella G. The bayesian lasso. J Am Stat Assoc. 2008;103(482):681–6. Association, vol. 50. New Zealand: New Zealand Grassland Association; 38. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32. 1989. p. 253–4. 39. Endelman JB. Ridge regression and other kernels for genomic selection 16. Collard BC, Mackill DJ. Marker-assisted selection: an approach for with R package rrBLUP. The Plant Genome. 2011;4(3):250–5. precision plant breeding in the twenty-first century. Phil Trans R Soc B Biol 40. de los Campos G, Pérez-Rodríguez P. Bayesian generalized linear Sci. 2008;363(1491):557–72. regression. R package version 1.0. 4. 2014. 17. Dumsday JL, Smith KF, Forster JW, Jones ES. SSR-based genetic linkage 41. Liaw A, Wiener M. Classification and regression by randomForest. R news. analysis of resistance to crown rust (Puccinia coronata f sp lolii) in 2002;2(3):18–22. perennial ryegrass (Lolium perenne). Plant Pathol. 2003;52(5):628–37. 42. Annicchiarico P, Nazzicari N, Li X, Wei Y, Pecetti L, Brummer EC. 18. Dracatos PM, Cogan NOI, Dobrowolski MP, Sawbridge TI, Spangenberg GC, Accuracy of genomic selection for alfalfa biomass yield in different Smith KF, Forster JW. Discovery and genetic mapping of single reference populations. BMC Genom. 2015;16(1):1020. nucleotide polymorphisms in candidate genes for pathogen defence 43. Crossa J, Beyene Y, Kassa S, Pérez P, Hickey JM, Chen C, De Los Campos G, response in perennial ryegrass (Lolium perenne L,). Theor Appl Genet. Burgueño J, Windhausen VS, Buckler E, et al. Genomic prediction in 2008;117(2):203–19. maize breeding populations with genotyping-by-sequencing. G3: Genes, 19. Dracatos PM, Cogan NO, Sawbridge TI, Gendall AR, Smith KF, Genomes, Genetics. 2013;3(11):1903–26. Spangenberg GC, Forster JW. Molecular characterisation and genetic mapping of candidate genes for qualitative disease resistance in 44. Schubiger FX, Baert J, Cagas B, Cernoch V, Chosson JF, Czembor E, perennial ryegrass (Lolium perenne L,). BMC Plant Biol. 2009;9(1):1. Eickmeyer F, Feuerstein U, Hartmann S, Jakesova H, et al. The eucarpia 20. Muylle H, Baert J, Van Bockstaele E, Moerkerke B, Goetghebeur E, multi-site rust evaluation–results 2007. In: Sustainable Use of Genetic Diversity Roldàn-Ruiz I. Identification of molecular markers linked with crown rust in Forage and Turf Breeding. New York: Springer; 2010. p. 331–40. Arojju et al. BMC Genetics (2018) 19:35 Page 10 of 10 45. Roderick HW, Thomas BJ. Infection of ryegrass by three rust fungi (Puccinia coronata, P graminis and P loliina) and some effects of temperature on the establishment of the disease and sporulation. Plant Pathol. 1997;46(5):751–61. 46. Roderick HW, Thorogood D, Adomako B. Temperature-dependent resistance to crown rust infection in perennial ryegrass, Lolium perenne. Plant Breed. 2000;119(1):93–5. 47. Schein RD, Rotem J. Temperature and humidity effects on uredospore viability. Mycologia. 1965;57(3):397–403. 48. Fè D, Pedersen MG, Jensen CS, Jensen J. Genetic and environmental variation in a commercial breeding program of perennial ryegrass. Crop Sci. 2015;55(2):631–40. 49. Muranty H, Troggio M, Sadok IB, Al Rifaï M, Auwerkerken A, Banchi E, Velasco R, Stevanato P, Van De Weg WE, Di Guardo M, et al. Accuracy and responses of genomic selection on key traits in apple breeding. Hortic Res. 2015;2:15060. 50. Wimmer V, Lehermeier C, Albrecht T, Auinger H-J, Wang Y, Schön CC. Genome-wide prediction of traits with different genetic architecture through efficient variable selection. Genetics. 2013;195(2):573–87. 51. Isidro J, Jannink J-L, Akdemir D, Poland J, Heslot N, Sorrells ME. Training set optimization under population structure in genomic selection. Theor Appl Genet. 2015;128(1):145–58. 52. Lorenz AJ, Smith KP, Jannink JL. Potential and optimization of genomic selection for fusarium head blight resistance in six-row barley. Crop Sci. 2012;52(4):1609–21. 53. Riedelsheimer C, Endelman JB, Stange M, Sorrells ME, Jannink J-L, Melchinger AE. Genomic predictability of interconnected biparental maize populations. Genetics. 2013;194(2):493–503. 54. Tayeh N, Klein A, Le Paslier MC, Jacquin F, Houtin H, Rond C, Chabert-Martinello M, Magnin-Robert JB, Marget P, Aubert G, et al. Genomic prediction in pea: effect of marker density and training population size and composition on prediction accuracy. Front Plant Sci. 2015;6:941. 55. Habier D, Fernando RL, Dekkers JCM. The impact of genetic relationship information on genome-assisted breeding values. Genetics. 2007;177(4):2389–97. 56. Zhang Z, Ding X, Liu J, Zhang Q, de Koning D-J. Accuracy of genomic prediction using low-density marker panels. J Dairy Sci. 2011;94(7): 3642–50. 57. Habier D, Fernando RL, Dekkers JCM. Genomic selection using low-density marker panels. Genetics. 2009;182(1):343–53. 58. Byrne SL, Nagy I, Pfeifer M, Armstead I, Swain S, Studer B, Mayer K, Campbell JD, Czaban A, Hentrup S, et al. A synteny-based draft genome sequence of the forage grass lolium perenne. Plant J. 2015;84(4):816–26. 59. Pfeifer M, Martis M, Asp T, Mayer KF, Lübberstedt T, Byrne S, Frei U, Studer B. The perennial ryegrass genomezipper: targeted use of genome resources for comparative grass genomics. Plant Physiol. 2013;161(2): 571–82. 60. Postel S, Küfner I, Beuter C, Mazzotta S, Schwedt A, Borlotti A, Halter T, Kemmerling B, Nürnberger T. The multifunctional leucine-rich repeat receptor kinase bak1 is implicated in arabidopsis development and immunity. Eur J Cell Biol. 2010;89(2):169–74. 61. Campbell NR, Harmon SA, Narum SR. Genotyping-in-thousands by sequencing (GT-seq): A cost effective SNP genotyping method based on custom amplicon sequencing. Mol Ecol Resour. 2015;15(4):855–67. 62. Proost S, Van Bel M, Vaneechoutte D, Van de Peer Y, Inzé D, Mueller-Roeber B, Vandepoele K. Plaza 30: an access point for plant comparative genomics. Nucleic Acids Res. 2015;43(D1):974–81. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BMC Genetics Springer Journals
Free
10 pages

Loading next page...
 
/lp/springer_journal/genomic-prediction-of-crown-rust-resistance-in-lolium-perenne-PyHziJHlJn
Publisher
Springer Journals
Copyright
Copyright © 2018 by The Author(s)
Subject
Life Sciences; Life Sciences, general; Animal Genetics and Genomics; Microbial Genetics and Genomics; Plant Genetics and Genomics; Genetics and Population Dynamics
eISSN
1471-2156
D.O.I.
10.1186/s12863-018-0613-z
Publisher site
See Article on Publisher Site

Abstract

Background: Genomic selection (GS) can accelerate genetic gains in breeding programmes by reducing the time it takes to complete a cycle of selection. Puccinia coronata f. sp lolli (crown rust) is one of the most widespread diseases of perennial ryegrass and can lead to reductions in yield, persistency and nutritional value. Here, we used a large perennial ryegrass population to assess the accuracy of using genome wide markers to predict crown rust resistance and to investigate the factors affecting predictive ability. Results: Using these data, predictive ability for crown rust resistance in the complete population reached a maximum of 0.52. Much of the predictive ability resulted from the ability of markers to capture genetic relationships among families within the training set, and reducing the marker density had little impact on predictive ability. Using permutation based variable importance measure and genome wide association studies (GWAS) to identify and rank markers enabled the identification of a small subset of SNPs that could achieve predictive abilities close to those achieved using the complete marker set. Conclusion: Using a GWAS to identify and rank markers enabled a small panel of markers to be identified that could achieve higher predictive ability than the same number of randomly selected markers, and predictive abilities close to those achieved with the entire marker set. This was particularly evident in a sub-population characterised by having on-average higher genome-wide linkage disequilibirum (LD). Higher predictive abilities with selected markers over random markers suggests they are in LD with QTL. Accuracy due to genetic relationships will decay rapidly over generations whereas accuracy due to LD will persist, which is advantageous for practical breeding applications. Keywords: Genomic selection, Crown rust, Perennial ryegrass, Genetic relationship, GWAS Background that can be utilized to develop resistant cultivars [11–13]. Perennial ryegrass (Lolium perenne L.) is the predominant Phenotypic recurrent selection is typically used to develop forage species grown in temperate regions of the world [1]. cultivars with improved resistance and selection is often Puccinia coronata f. sp. lolli (crown rust) is one of the most carried out on spaced plants [9, 11, 12, 14]. There is a high widespread diseases of perennial ryegrass and can lead to correlation between spaced plants and swards for disease a reduction in forage nutritive value, yield and persistency resistance and indirect selection for disease resistance [2–4]. Poor quality, rust infected swards can impact ani- on spaced plants can improve resistance in sward condi- mal performance and well-being [5–7]. Developing resis- tions [15]. However, with the advancements in molecu- tant cultivars is the most viable option for disease control lar marker development over the last decade, efforts to and it has been shown that resistance to crown rust is use marker assisted breeding strategies have been pur- conferred by both quantitative and qualitative inheritance sued. One such strategy involves identifying quantitative [8–11]. As an obligate out-crossing species, perennial rye- trait loci (QTL) in bi-parental mapping populations and grass germplasm has high variation for disease resistance using markers to efficiently backcross the QTL into elite breeding material [16]. Although QTLs explaining signif- *Correspondence: stephen.byrne@teagasc.ie icant phenotypic variation for crown rust resistance were Teagasc, Crop Science Department, Oak Park, R93 XE12 Carlow, Ireland mapped onto linkage group (LG) 1-5 and 7 [17–23], it is Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Arojju et al. BMC Genetics (2018) 19:35 Page 2 of 10 unclear if any of these QTLs were successfully introduced a glasshouse and later transplanted to the field in 2013 at into breeding material. Genome wide association studies Oak Park, Carlow, Ireland (52° 51 34.2 N,6°55 03.0 W). (GWAS) are another approach to identify markers linked Plants were grown in two replicates in a partially balanced to QTL. In this case breeding populations can directly be incomplete block design. Each block consists of 60 test used to identify marker-trait associations, although iden- genotypes and 5 check genotypes and was surrounded tified markers tended to explain a small proportion of the by a 1.5 m sward consisting of a four way mix of crown total additive genetic variance, resulting in smaller genetic rust susceptible perennial ryegrass cultivars. Crown rust gains [24–26]. was recorded in the years 2014 and 2015 as mean per- Genomic selection (GS) was first proposed by Meuwissen centage disease score on each plant. Briefly, percentage et al. [27], as a method to capture complete additive disease score was obtained by combining scores of per- genetic variance using genome wide markers. GS is a form centage of leaves with infection and average percentage of of marker assisted breeding, which accounts for all marker infection on diseased leaves. Scoring was done at multi- effects across the entire genome to calculate genomic esti- ple time points in September to November without any mated breeding values (GEBVs), which are used to select harvest cuts between scorings (Table 1). We are trying individual plants for advancement [26]. Use of genome- to develop genomic models to identify plants with good wide markers will include small effect loci and is ideal resistance to crown rust across the season, and we decided for complex traits with low to moderate heritability. In to use all time points for constructing a quantitative sum- GS, a training population is genotyped with genome wide mary for crown rust resistance. To do this we calculated markers and phenotyped for the trait under selection AUDPC for each spaced plant in both years. Using mul- and models to predict breeding values from marker data tiple time point data, we generated area under disease are developed. Implementing GS for complex traits like progress curve (AUDPC) as follows: yield and quality is a primary objective of many peren- N −1 (y + y ) i i+1 nial ryegrass breeding programmes. In contrast to yield A = (t − t ) (1) k i+1 i and quality traits, the cost (labour and time) of pheno- i=1 typing for disease resistance is much lower. However, it where y is the extent of infection (percentage disease is important that any GS approaches targeting yield and th score) at i observation and t isthetimepoint at quality improvements also ensure adequate disease resis- th i observation. N is the number of individuals in the tance is maintained, particularly where multiple rounds data set. of marker based selections are performed between field Variance components for crown rust were estimated evaluations. Opportunities for GS in perennial ryegrass using the R package lme4 (linear mixed-effects models were first reviewed by Hayes et al. [28], and the earli- using ’eigen’ and S4) [34]. Broad sense heritability was est empirical study was done by Fè et al. [29] for heading estimated as follows: date, which confirmed the superiority of GS over marker assisted selection. Later Fè et al. [30], Grinberg et al. [31] g H = (2) and Byrne et al. [32] reported high predictive ability for 2 2 2 σ + σ /2 + σ /4 g g∗yr res important agronomical traits in perennial ryegrass. In particular, predictive ability for crown rust reached up to where σ is the total genetic variance among individu- 0.58 [30] when genotypes and phenotypes were evaluated als, σ is the variance associated with genotype by year g∗yr on F families. In this study, we evaluated predictive abil- ity for crown rust resistance on individual plants in a large Table 1 Mean percentage disease score for crown rust resistance perennial ryegrass population, and assessed factors con- at different time points (TP) in Year1 (2014) and Year2 (2015) tributing to predictive ability, such as training population Time point/dates Mean SD Min Max size and marker density. We also performed a GWAS to identify a small to moderately sized panel of markers with Year 1 good predictive ability for crown rust resistance. TP1 (13/10/14) 3.1 6.1 0 40 TP2 (20/10/14) 5.2 7.6 0 45 Methods TP3 (29/10/14) 9.6 10.8 0 60 Plant material, phenotyping and genotyping TP4 (10/11/14) 9.8 8.7 0 45 The training population consists of 30 diploid peren- nial ryegrass families that have been described previously Year 2 [32, 33]. Each family consists of 60 genotypes making up a TP1 (21/09/15) 2.0 4.4 0 32 population of 1800 individuals. The complete population TP2 (05/10/15) 11.2 10.0 0 60 consists of ten cultivars, eight full-sib families, eight half- TP3 (19/10/15) 19.9 9.0 0 63 sib families and four ecotypes. Plants were established in Arojju et al. BMC Genetics (2018) 19:35 Page 3 of 10 interaction and σ is residual variance. With genotype as repeated 100 times and the resulting accuracies were then res random effect and year and checks as fixed effects, condi- averaged. This approach to cross-validation has previ- tional modes (BLUPs) were calculated in lme4 and used as ously been used to evaluate genomic prediction mod- input for genomic prediction. els [42, 43]. Predictive ability and bias were assessed Genotyping was done using genotyping by sequenc- in the complete population and in each sub-population. ing (GBS) approach described by Elshire et al. [35]and Predictive ability (r ) was determined as the Pearson’s data analysed as described in Byrne et al. [32]. Briefly, correlation coefficient between observed phenotypic genomic DNA was extracted from leaf samples and GBS value and predicted phenotype. Bias was evaluated by libraries were prepared using the restriction enzyme regressing observed phenotypic value on predictions. ApeKI, libraries were amplified and sequenced on an Illu- We reduced training population size and marker den- mina Hiseq2000. Panels of SNPs were identified in the sity in order to identify the impact of training population complete population, as well as in all sub populations size and marker number on predictive ability. To com- separately (half-sibs, full-sibs, ecotypes, cultivars). Indi- pare predictive ability for traits with contrasting genetic viduals with very low sequencing coverage and/or largely architecture we compared heading date, a highly heri- missing phenotypic data were eliminated from the anal- table trait, with crown rust. Predictive ability for head- ysis giving a final population for analysis of 1582 indi- ing date has already been shown to be high (0.81) in viduals. Missing marker data was imputed using mean this population [32]. We re-analyzed data for heading imputation. date according to methods described above and made a comparison with crown rust. To evaluate the impact of Genomic prediction models leaving related material out of the training set we also We used four statistical algorithms for genomic pre- performed cross validation by leaving one family out. In diction, ridge regression best linear unbiased prediction this approach one complete family (up to 60 individuals) (rrBLUP) [27], Bayes B [36] and Bayesian Lasso [37], and is left out of the training set and only used for testing. random forest [38]. rrBLUP is a mixed model approach, This was repeated so that each family in turn is used as a which was initially proposed for GS. We used an R pack- test set. age called rrBLUP [39] for fitting the mixed model as follows Genome wide association A mixed linear model (MLM) was also used for associ- y = μ + Xg +  (3) ation mapping, implemented in the R package rrBLUP where μ istheoverall mean,Xisthemarkermatrix, gis [39]. Population structure and family relatedness was accounted for with a kinship matrix calculated by rrBLUP the matrix of marker effects,  is a vector of residual from the input genotypic data. We accounted for multiple effects and y is a vector of conditional modes for crown testing using a Bonferroni correction and markers pass- rust. We also evaluated two Bayesian approaches, Bayes B ing an α level 0.05 threshold were considered statistically [36] and Bayesian Lasso [37], which were both imple- significant. mented using the R package BGLR [40] with the following parameters: number of iterations = 5000, burn-in = 500 and thinning = 5. Random forest is a machine-learning Results and discussion tool, in which series of regression trees were grown inde- Phenotypic analysis for crown rust pendently to the largest extent possible using subsets of The mean percentage disease score for crown rust infec- bootstrap samples. At each split of the tree, a random tion in the population increased over time in both eval- subset of variables is selected to identify the best split. uation years as infection levels accumulated (Table 1). In We implemented random forest using the R package ran- both years, evaluations were carried out in the period domForest [41], setting the number of variables at each from September to November during a time when disease split to 1/3 of the total variables, and using a terminal pressure tends to be at its greatest [15, 44]. The highest node size of five and minimum of 500 trees per forest. We mean percentage disease score was seen in late October also used random forest to rank variables using the vari- 2015 and was more than double the highest mean percent- able importance measure, a permutation based measure age disease score from 2014 (Table 1). In addition to plant in which variables are ranked based on the mean decrease health and level of host resistance, crown rust infection is in accuracy. influenced by various environmental factors, such as tem- perature, relative humidity, and light [45–47]. The latency Cross validation scheme period is reduced and spore production increased as tem- We evaluated genomic prediction models using Monte- perature increases [45], and it has been shown that when Carlo cross-validation by randomly assigning plants into temperatures exceed 25°C, the susceptibility of previously training (70%) and test (30%) sets and the procedure was resistant cultivars can be increased [46]. It has already Arojju et al. BMC Genetics (2018) 19:35 Page 4 of 10 been shown that there is variability within pathogen pop- and used to aid selection of the top performing families ulations, and different races can be found within and from which to construct the synthetics cultivars. During between locations. It is also possible that the composition construction of synthetics a spaced plant nursery may be of a pathogen population can change over short periods of established to evaluate heading date and crown rust resis- time and plants that are resistant at one point in time will tance before selecting individual genotypes from which to become susceptible as the pathogen population shifts or construct synthetics (within family selection). In practice, evolves. this has a time cost of 2 to 3 years (establishment, eval- AUDPC values ranged from 0 to 1371 and the Pearson uation, selection and recombining), and using molecular correlation co-efficient between replicates within years markers offers an opportunity to reduce this to one year was moderate (0.69 in 2014 and 0.59 in 2015). However, in those selection cycles where GEBVs are predicted. This the Pearson correlation co-efficient between years was depends on our ability to accurately predict traits such as low (0.28), and there was a significant genotype by year crown rust from genomic data. interaction (F = 3.025, MSE = 60676, p = 0.0001). (1761) The broad sense heritability for crown rust infection Predicting crown rust resistance with genomic data was moderate (0.36), which is in line with previous We evaluated four algorithms for prediction of crown rust estimates of heritability calculated in other populations infection from genomic data, rrBLUP, Bayes B, Bayesian [11, 48]. Overall there is a good phenotypic variation Lasso, and random forest. The mean predictive ability for crown rust infection among and within the 30 fam- after cross-validation within the complete population was ilies/cultivars/ecotypes making up the entire population 0.52 using rrBLUP, 0.52 using Bayesian Lasso, 0.51 using (Fig. 1). Plants were placed into one of four categories Bayes B, and 0.49 using random forest (Additional file 1: (sub-populations) based on mating type or origin, these Figure S1). rrBLUP was computationally faster, and there- were (i) full-sib families, (ii) half-sib families, (iii) cultivars, fore results from all further analysis are only reported and (iv) ecotypes. In general the ecotypes were more sus- for rrBLUP. The predictive ability of 0.52 is in line with ceptible to crown rust infection than cultivars or breed- previous estimates reported in perennial ryegrass where ing material (Fig. 1), which presents a challenge for the predictions were based on mean genotypes and pheno- incorporation of ecotypes into breeding programmes. The types of F families [30]. Predictive ability did not dif- broad-sense heritability calculated in each sub-population fer depending on whether the equations were developed varied between 0.17 in the cultivars to 0.44 in the full-sib using phenotypes from the last time point scored or the families. AUDPC values incorporating all time points. This indi- Crown rust infection is typically evaluated in breeding cates that a single scoring each year would have sufficed. programmes by growing spaced plants or potted plants However,theimportanceofevaluatingcrown rust in more from a population and visually scoring the level of crown than one year was emphasised by the low correlation rust infection. A mean score is assigned to each family between scores in 2014 and 2015. When we calculate the predictive ability within each of the sub-populations (cultivars, half-sib families, full-sib families, and ecotypes), the highest predictive ability for crown rust was obtained using plants from full-sib fami- lies (0.54) and the lowest predictive ability for crown rust was obtained with the plants from the ecotypes (0.24) (Fig. 2). Generally, traits with higher heritability achieve higher predictive abilities [49, 50], and we see that here where crown rust measurements taken in the full-sib families had the highest broad-sense heritability and the highest predictive ability. In general, there was a good cor- relation between predictive ability and both phenotypic variance and heritability. This relationship between phe- notypic variance and predictive ability has been observed previously [49, 51]. We also evaluated the predictive ability using a leave- one-family-out cross validation scheme. The complete Fig. 1 Phenotypic variation for crown rust resistance in the complete population, grouped according to sub-population types: cultivars population is comprised of 30 families/cultivars/ecotypes, (CS), ecotypes (ES), full-sibs (FS) and half-sibs (HS). Broad sense each with up to 60 individual genotypes. The predic- heritability (H ) in complete population and sub-populations is tive ability was assessed in the complete population by highlighted over the figure selectively leaving one family out of the training set and Arojju et al. BMC Genetics (2018) 19:35 Page 5 of 10 Fig. 2 Predictive ability in different population types. Complete population (CP), cultivars (CS), ecotypes (ES), full-sibs (FS) and half-sibs (HS) are listed on x-axis, predictive ability (left) and bias (right) on y-axis. Crown rust is in red and heading date in blue using it for testing. In addition to crown rust we also The drop in predictive ability was more pronounced as we evaluated predictive ability for heading date phenotypes reduced the training population size for crown rust resis- previously reported [32]. The predictive ability for both tance than it was for heading date. The predictive ability crown rust (r = 0.02, min =−0.36, max = 0.36) and for crown rust resistance when using 90% of the popula- heading date (r = 0.29, min =−0.14, max = 0.65) varied tion as a training set was 0.52 and the predictive ability was greatly depending on which family was left out, and hav- 0.38 when using just 10% of the population. Irrespective ing related material in the training set (shared parentage) of the trait, as the training population size increased there greatly improved predictive ability. was an increase in predictive ability which is consistent with similar correlations between training population size Effect of training population size and marker density on and predictive ability reported previously for perennial predictive ability ryegrass [29, 30]and othercrops [51–54]. Useful linkage As we reduced the number of individuals in the train- disequilibrium (LD) only extends over short distances in ing population we saw a decrease in predictive ability and perennial ryegrass and it has been suggested that this is an increasingly upward bias in the variance of predictions the result of a very large past effective population size [28]. This impacts both the size of the reference population for both crown rust resistance and heading date (Fig. 3). Fig. 3 Effect of training population size on predictive ability. Training population is varied from 90% (1423 individuals) to 10% (158 individuals) on x-axis and predictive ability (left), bias (right) on y-axis. Crown rust is in red and heading date in blue Arojju et al. BMC Genetics (2018) 19:35 Page 6 of 10 and marker density required to achieve high accuracies Identifying SNPs associated with crown rust resistance when predicting traits from genomic data. The fact that The cost of genotyping impacts the number of selection we are able to achieve high predictive abilities with rela- candidates that can be evaluated and therefore impacts tively small training populations is likely a result of strong the selection intensity. Different approaches to low den- genetic structure and differentiation in our diverse popu- sity SNP genotyping for genomic selection have been lation and the use of the marker data to capture genetic proposed. These include variable selection methods to relationships [55]. identify a small subset of markers in strong LD with the The limited LD also affects the number of markers trait [56] or using a small random subset of markers to required to obtain high predictive accuracies, and given impute from low-to-high density [57]. Until a chromo- the extent of LD in the broader perennial ryegrass pop- some scale assembly of the perennial ryegrass genome ulation, marker numbers in excess of one million have becomes available the latter remains a challenge. We used been suggested for achieving high accuracies [28]. When both permutation based variable importance measures we reduced marker number in the complete population and GWAS analysis to identify a subset of markers capa- and the various sub-populations we observed very little ble of predicting crown rust resistance. Using permutation impact on the predictive ability for either trait (Table 2). basedvariableimportancemeasureswewereabletorank Reducing the marker set to 5% of the total available had markers by mean decrease in accuracy and select the top virtually no impact on predictive ability in all cases. This ranked markers for use in genomic prediction. In the would support our observation that much of the predic- case of GWAS we ranked SNPs based on significance and tive ability can be derived from makers capturing familial again selected the top ranked markers for use in genomic relationships. When marker number dropped below 5% prediction. All variable importance measures and GWAS (10878) predictive ability for both traits in the complete were identified and ranked in the training set and used population began to drop. However, even with 0.05% to predict phenotypes in the test set via cross-validation. (109) of markers the mean predictive ability was 0.30 for When we used the top 100 ranked markers from the per- crown rust resistance and 0.52 for heading date. Know- mutation based variable importance measures, the mean ing the contribution of genetic relationships to predictive predictive ability of 100 iterations was 0.42 (ranging from ability is important because it will change over genera- 0.36 to 0.48). When we used the top 100 ranked markers tions. In contrast, predictive ability due to LD has greater from the GWAS analysis, the mean predictive ability of persistence over generations and is therefore preferen- 100 iterations was 0.36 (ranging from 0.25 to 0.44). In both tial [55]. Schemes for implementing genomic selection cases the mean predictive ability with selected markers in perennial ryegrass that pursue a reduction in effec- is higher than the predictive ability with random mark- tive population size from the outset have been proposed. ers, which was 0.28 (ranging from 0.18 to 0.39). The lower Such schemes would lead to an increase in the extent predictive ability using GWAS marker selection is not of LD and ensure that predictive ability due to LD can surprising considering that we corrected for population be captured using a reasonable number of markers and structure using a kinship matrix, and we are more reliant a reference population size that is feasible in breeding on identifying markers in LD with the trait. As discussed programmes. above, the predictive ability of these markers is expected Table 2 Predictive ability (r ) and bias for crown rust (CR) and heading date (HD) by selecting random markers of 100 to 0.05%, in complete population (CP), cultivars (CS), full-sibs (FS) and half-sibs (HS) Pop 100% 60% 20% 5% 1% 0.5% 0.1% 0.05% r bias r bias r bias r bias r bias r bias r bias r bias p p p p p p p p CR CP 0.52 1.22 0.52 1.22 0.52 1.21 0.51 1.18 0.46 1.10 0.43 1.07 0.36 1.04 0.30 1.48 CS 0.29 1.28 0.28 1.26 0.28 1.24 0.27 1.18 0.22 0.97 0.17 0.80 0.14 0.95 0.10 1.13 FS 0.54 1.13 0.54 1.13 0.54 1.13 0.54 1.14 0.52 1.07 0.50 1.03 0.45 1.00 0.40 0.99 HS 0.49 1.24 0.49 1.24 0.49 1.24 0.49 1.24 0.48 1.23 0.46 1.21 0.42 1.23 0.36 1.22 HD CP 0.81 1.16 0.81 1.16 0.81 1.16 0.80 1.14 0.75 1.07 0.72 1.05 0.62 1.01 0.52 1.00 CS 0.84 1.25 0.81 1.19 0.81 1.20 0.81 1.18 0.78 1.11 0.77 1.12 0.66 1.03 0.56 1.02 FS 0.76 1.00 0.75 1.16 0.75 1.16 0.75 1.16 0.74 1.14 0.68 1.27 0.64 1.26 0.62 0.87 HS 0.74 1.18 0.74 1.09 0.74 1.10 0.74 1.09 0.73 1.08 0.72 1.15 0.67 1.10 0.62 1.09 Arojju et al. BMC Genetics (2018) 19:35 Page 7 of 10 to be more persistent over subsequent generations. Using All markers were located within 16 genomic scaffolds GWAS selectedmarkersitisclear to seethattheyare containing 44 predicted genes. Out of 16 scaffolds we superior to randomly selected markers up to the point, were able to place seven scaffolds onto LG3, 5 and 7 beyond which adding more markers does not improve (Additional file 3: Table S2). We found five common scaf- predictive ability in either case (Fig. 4). The ability of a folds between the complete population and the IBERS GWAS within each sub-population to identify and select and only two of these scaffolds were mapped, onto LG3. a small set of SNPs with excellent predictive ability var- On LG3 five markers were anchored within 60.4-61.21 ied, and in some cases was little better than random SNP cM. Genes present on these scaffolds were coding for selection (Fig. 5). The GWAS on plants originating from domains including Mon1, Aquaporin, DUF1635, Nucleo- IBERs bred cultivars identified a small set of twenty SNPs redoxin, Beta-glucan export ATP-binding/permease pro- with 77% of the predictive ability achieved with 20,000 tein, BRASSINOSTEROID INSENSITIVE 1-associated SNPs. The power of a GWAS to identify markers with receptor kinase 1, Alpha N-terminal protein methyltrans- high predictive ability was much greater within the popu- ferase 1. Gene function of these domains plays a key role in lation made up of IBERs plants than within cultivars, and ATP-binding, membrane proteins, enzyme catalysis and full-sib families where twenty SNPs could only achieve 46 pathogen-associated molecular pattern (PAMP)-triggered and 48% of the predictive ability with 20,000 SNPs, respec- immunity (PTI) (Additional file 4:TableS3)[60]. tively. On average LD is higher within the sub-population Using small subsets of trait associated markers may be with IBERs plants, which may explain the greater an effective strategy for within-family prediction of traits ability to identify markers associated with crown-rust such as heading date, crown rust resistance and some resistance. quality traits. Predicting heading date from markers would In order to characterise the markers associated with enable plants to be matched in heading date to ensure crown rust resistance we repeated the GWAS analysis sufficient cross-pollination when constructing synthetic without division of genotypes into training and testing cultivars [32]. Combining these with markers to predict sets. We carried out GWAS using the complete pop- crown rust resistance would also avoid the inclusion of ulation and found 29 markers significantly associated plants with high levels of susceptibility, and furthermore with crown rust resistance after correction for multiple prediction models can be based on multi-year evaluations. testing (Additional file 2: Table S1). Using the peren- It is clear from the phenotypic data presented here that nial ryegrass genome [58] as a reference, we located all there is substantial within family variation for crown rust markers within 22 genomic scaffolds that contained 50 resistance. Opportunities already exist to genotype small predicted genes. Using the Genome Zippper [58, 59], to moderate sized marker panels in 1000s of samples at we anchored ten scaffolds onto LG2, 3, 4, 5 and 7 low cost [61]. Using these approaches small fragments (Additional file 3: Table S2). Similarly, we did GWAS (200-300 bp) are amplified and sequenced at hundreds of on IBERS material and found 24 markers associated loci. These amplicons can be used as short haplotypes in with crown rust resistance (Additional file 2:TableS1). marker aided selection strategies. An assay can be devel- oped to target loci in linkage with QTL for heading date [32], crown rust resistance, and other traits such as qual- ity parameters. The assay can also include a suite of loci with a good distribution throughout the genome and be deployed for among-and-within-full-sib-family selection (Additional file 5: Figure S2). Once high yielding fami- lies are identified in field trials, within family selection for crown rust resistance and forage quality can be per- formed at a high selection intensity with the molecular marker assay. Furthermore, plants can be selected to be synchronous in flowering time. Conclusions Our findings show that predicting crown rust resistance in perennial ryegrass can be achieved with high accuracy using AUDPC scores on spaced plants. However, there was no difference in predictive ability when equations Fig. 4 Predictive ability of selected markers versus random markers in the complete population. Markers were selected based on the were developed using phenotypes from the last time-point ranking from genome wide association studies and compared with scored or the AUDPC values, meaning a single time point random markers of similar size was adequate to evaluate the crown rust susceptibility of Arojju et al. BMC Genetics (2018) 19:35 Page 8 of 10 Fig. 5 Comparing predictive ability of selected versus random markers. Markers were selected based on the ranking from genome wide association studies in cultivars, full-sibs and IBERS material and compared with random markers of similar size the spaced plants. Much of the predictive ability comes Abbreviations AUDPC: Area under disease progress curve; GEBV: Genomic estimated from markers capturing familial relationships, highlighted breeding value; GS: Genomic selection; GWAS: Genome wide association by the observation that there was no drop in predic- study; LD: Linkage disequilibrium; QTL: Quantitative trait locus tive ability when going from the entire marker set down Acknowledgments to only 5% (10,878) of the marker set. Accuracy due Authors would like to thank Jean-Baptiste Enjelvin, Olivia Aylesbury, Mary to genetic relationships will decay rapidly over genera- O’Sullivan, Padraig Dunne and Michael Murphy for help with phenotyping and Sean Murray and Helena Meally for genomic DNA extraction. tions whereas accuracy due to LD will persist. Using a GWAS we attempted to identify and rank markers in Funding This work received funding from the Irish Department of Agriculture Food and LD with QTL. This enabled a small panel of markers to the Marine DAFM (RSF 11/S/109) and Teagasc core funding. SKA is supported be identified that had higher predictive ability than the by a Teagasc PhD Walsh Fellowship. SLB has received funding from the same number of randomly selected markers, and had pre- European Union’s Horizon 2020 research and innovation programme under dictive abilities close to those achieved with the entire the Marie Sklodowska-Curie grant agreement No. 658031. The funding bodies played no role in the design of the study and collection, analysis, and marker set. interpretation of data and in writing the manuscript. Availability of data and materials Additional files All sequence data associated with this project has been deposited in the NCBI Short Read Archive (SRA) under the BioProject accession PRJNA352789. Phenotypic data is available on Figshare (https://doi.org/10.6084/m9.figshare. Additional file 1: Figure S1. Predictive ability and bias for crown rust 5562640.v1), along with an R script template used for cross-validation (https:// using various algorithms for genomic prediction. (PDF 101 kb) doi.org/10.6084/m9.figshare.5712208.v1). Additional file 2: Table S1. List of markers associated with crown rust resistance based on genome wide association studies in complete Authors’ contributions population and IBERS material. (XLSX 14 kb) DM, PC, SB, TRH, MC conceived and designed the study. SKA and SLB performed the data analysis and drafted the initial manuscript. SKA, SLB, DM, Additional file 3: Table S2. List of genomic scaffolds where all the PC, TRH, MC, TM and SB contributed to interpretation of data and preparation significant markers from genome wide association studies were located. of the final manuscript. All authors read and approved the final version. Scaffolds were placed onto linkage group with the aid of Genome Zipper [59]. (XLSX 11 kb) Ethics approval and consent to participate Additional file 4: Table S3. List of predicted proteins on the genomic Not applicable. scaffolds. Markers located on these scaffolds were associated with crown rust resistance. BLAST was done on the predicted protein sequences using Competing interests PLAZA [62] to obtain the gene function. (XLSX 17 kb) The authors declare that they have no competing interests. Additional file 5: Figure S2. Among-and-within-full-sib-family selection Publisher’s Note that incorporates an inexpensive genotyping assay to implement Springer Nature remains neutral with regard to jurisdictional claims in within-family selection using a high selection intensity. (PDF 227 kb) published maps and institutional affiliations. Arojju et al. BMC Genetics (2018) 19:35 Page 9 of 10 Author details (Puccinia coronata f sp lolii) resistance in perennial ryegrass (Lolium Teagasc, Crop Science Department, Oak Park, R93 XE12 Carlow, Ireland. perenne) using AFLP markers and a bulked segregant approach. Teagasc, Grassland Science Research Department, Animal and Grassland Euphytica. 2005;143(1-2):135–44. Research and Innovation Centre, Oak Park, R93 XE12 Carlow, Ireland. 21. Muylle H, Baert J, Van Bockstaele E, Pertijs J, Roldàn-Ruiz I. Four QTLs Department of Agronomy, University of Wisconsin-Madison, WI53706 determine crown rust (Puccinia coronata f sp lolii) resistance in a perennial Madison, USA. Agricultural Research Service, United State Department of ryegrass (Lolium perenne) population. Heredity. 2005;95(5):348–57. Agriculture, WI53706 Madison, USA. Department of Botany, School of Natural 22. Schejbel B, Jensen LB, Xing Y, Lübberstedt T. QTL analysis of crown rust Sciences, Trinity College Dublin, Dublin 2 Ireland. resistance in perennial ryegrass under conditions of natural and artificial infection. Plant Breed. 2007;126(4):347–52. Received: 16 June 2017 Accepted: 18 April 2018 23. Thorogood D, Paget MF, Humphreys MO, Turner LB, Armstead I, Roderick H. QTL analysis of crown rust resistance in perennial ryegrass-implications for breeding. Int Turfgrass Soc Res J. 2001;9:218–23. 24. Hayes B, Goddard M. Genome-wide association and genomic selection References in animal breeding this article is one of a selection of papers from the 1. Humphreys M, Feuerstein U, Vandewalle M, Baert J. Ryegrasses. conference “exploiting genome-wide association in oilseed brassicas: a In: Fodder Crops and Amenity Grasses. New York: Springer; 2010. p. 211–60. model for genetic improvement of major OECD crops for sustainable 2. Potter LR. Effect of crown rust on regrowth, competitive ability and farming”. Genome. 2010;53(11):876–83. nutritional quality of perennial and Italian ryegrasses. Plant Pathol. 25. Heffner EL, Sorrells ME, Jannink J-L. Genomic selection for crop 1987;36(4):455–61. improvement. Crop Sci. 2009;49(1):1–12. 3. Valk P, Proveniers MCG, Pertijs JH, Lamers JTW, Dun CMP, Smeekens JCM. Late heading of perennial ryegrass caused by introducing an Arabidopsis 26. Jannink J-L, Lorenz AJ, Iwata H. Genomic selection in plant breeding: homeobox gene. Plant Breed. 2004;123(6):531–5. from theory to practice. Brief Funct Genom. 2010;9(2):166–77. 4. Price T. Ryegrass rust in victoria. Plant Protect Quart. 1987;2:189. 27. Meuwissen TH, Hayes BJ, Goddard ME. Prediction of total genetic value 5. Lancashire JA, Latch GCM. The effect of crown rust (Puccinia coronata using genome-wide dense marker maps. Genetics. 2001;157(4):1819–29. Corda) on the yield and botanical composition of two ryegrass/white 28. Hayes BJ, Cogan NO, Pembleton LW, Goddard ME, Wang J, clover pastures. N Z J Agric Res. 1970;13(2):279–86. Spangenberg GC, Forster JW. Prospects for genomic selection in forage 6. Plummer RM, Hall RL, Watt TA. The influence of crown rust (Puccinia plant species. Plant Breed. 2013;132(2):133–43. coronata) on tiller production and survival of perennial ryegrass (Lolium 29. Fè D, Cericola F, Byrne S, Lenk I, Ashraf BH, Pedersen MG, Roulund N, perenne) plants in simulated swards. Grass Forage Sci. 1990;45(1):9–16. Asp T, Janss L, Jensen CS, et al. Genomic dissection and prediction of 7. Smit HJ, Tas BM, Taweel HZ, Tamminga S, Elgersma A. Effects of heading date in perennial ryegrass. BMC Genomics. 2015;16(1):1. perennial ryegrass (Lolium perenne L,) cultivars on herbage production, 30. Fè D, Ashraf BH, Pedersen MG, Janss L, Byrne S, Roulund N, Lenk I, nutritional quality and herbage intake of grazing dairy cows. Grass Forage Didion T, Asp T, Jensen CS, et al. Accuracy of genomic prediction in a Sci. 2005;60(3):297–309. commercial perennial ryegrass breeding program. Plant Genome. 8. Hayward MD. Genetic control of resistance to crown rust (Puccinia 2016;9(3). coronato Corda) in Lolium perenne L, and its implications in breeding. 31. Grinberg NF, Lovatt A, Hegarty M, Lovatt A, Skøt KP, Kelly R, Blackmore T, Theor Appl Genet. 1977;51(2):49–53. Thorogood D, King RD, Armstead I, et al. Implementation of genomic 9. Kimbeng CA. Genetic basis of crown rust resistance in perennial ryegrass, prediction in Lolium perenne (L) breeding populations. Front Plant Sci. breeding strategies, and genetic variation among pathogen populations: 2016;7(133). a review. Anim Prod Sci. 1999;39(3):361–78. 32. Byrne S, Conaghan P, Barth S, Arojju SK, Casler M, Thibauld M, 10. McVeigh K. Breeding for resistance to crown rust (Puccinia coronata Velmurugan J, Milbourne D. Using variable importance measures to Corda var lolii Brown) in turf-type perennial ryegrass (Lolium perenne). identify a small set of SNPs to predict heading date in perennial ryegrass. New Brunswick: Rutgers University; 1975. PhD thesis. Sci Rep. 2017;7(3566). 11. Reheul D, Ghesquiere A. Breeding perennial ryegrass with better crown 33. Arojju SK, Barth S, Milbourne D, Conaghan P, Velmurugan J, Hodkinson TR, rust resistance. Plant Breed. 1996;115(6):465–9. Byrne SL. Markers associated with heading and aftermath heading in 12. Gibbs JG. Field resistance in Lolium sp to leaf rust (Puccinia coronata). perennial ryegrass full-sib families. BMC Plant Biol. 2016;16(1):160. Nature. 1966;209(5021):420. 34. Bates D, Mächler M, Bolker B, Walker S. Fitting linear mixed-effects 13. McGrath S, Hodkinson TR, Charles TM, Zen D, Barth S. Variation in models using lme4. arXiv preprint arXiv:14065823. 2014. inflorescence characters and inflorescence development in ecotypes and 35. Elshire RJ, Glaubitz JC, Sun Q, Poland JA, Kawamoto K, Buckler ES, cultivars of lolium perenne l. Grass Forage Sci. 2010;65(4):398–409. Mitchell SE. A robust, simple genotyping-by-sequencing (GBS) approach 14. Wilkins PW. Specialisation of crown rust on highly and moderately for high diversity species. PLoS ONE. 2011;6(5):19379. resistant plants of perennial ryegrass. Ann Appl Biol. 1978;88(1):179–84. 36. Habier D, Fernando RL, Kizilkaya K, Garrick DJ. Extension of the bayesian 15. Easton HS, Cooper BM, Frasers TJ, Widdup KH. Crown rust on perennial alphabet for genomic selection. BMC Bioinformatics. 2011;12(1):186. ryegrass in field trials. In: Proceedings of the New Zealand Grassland 37. Park T, Casella G. The bayesian lasso. J Am Stat Assoc. 2008;103(482):681–6. Association, vol. 50. New Zealand: New Zealand Grassland Association; 38. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32. 1989. p. 253–4. 39. Endelman JB. Ridge regression and other kernels for genomic selection 16. Collard BC, Mackill DJ. Marker-assisted selection: an approach for with R package rrBLUP. The Plant Genome. 2011;4(3):250–5. precision plant breeding in the twenty-first century. Phil Trans R Soc B Biol 40. de los Campos G, Pérez-Rodríguez P. Bayesian generalized linear Sci. 2008;363(1491):557–72. regression. R package version 1.0. 4. 2014. 17. Dumsday JL, Smith KF, Forster JW, Jones ES. SSR-based genetic linkage 41. Liaw A, Wiener M. Classification and regression by randomForest. R news. analysis of resistance to crown rust (Puccinia coronata f sp lolii) in 2002;2(3):18–22. perennial ryegrass (Lolium perenne). Plant Pathol. 2003;52(5):628–37. 42. Annicchiarico P, Nazzicari N, Li X, Wei Y, Pecetti L, Brummer EC. 18. Dracatos PM, Cogan NOI, Dobrowolski MP, Sawbridge TI, Spangenberg GC, Accuracy of genomic selection for alfalfa biomass yield in different Smith KF, Forster JW. Discovery and genetic mapping of single reference populations. BMC Genom. 2015;16(1):1020. nucleotide polymorphisms in candidate genes for pathogen defence 43. Crossa J, Beyene Y, Kassa S, Pérez P, Hickey JM, Chen C, De Los Campos G, response in perennial ryegrass (Lolium perenne L,). Theor Appl Genet. Burgueño J, Windhausen VS, Buckler E, et al. Genomic prediction in 2008;117(2):203–19. maize breeding populations with genotyping-by-sequencing. G3: Genes, 19. Dracatos PM, Cogan NO, Sawbridge TI, Gendall AR, Smith KF, Genomes, Genetics. 2013;3(11):1903–26. Spangenberg GC, Forster JW. Molecular characterisation and genetic mapping of candidate genes for qualitative disease resistance in 44. Schubiger FX, Baert J, Cagas B, Cernoch V, Chosson JF, Czembor E, perennial ryegrass (Lolium perenne L,). BMC Plant Biol. 2009;9(1):1. Eickmeyer F, Feuerstein U, Hartmann S, Jakesova H, et al. The eucarpia 20. Muylle H, Baert J, Van Bockstaele E, Moerkerke B, Goetghebeur E, multi-site rust evaluation–results 2007. In: Sustainable Use of Genetic Diversity Roldàn-Ruiz I. Identification of molecular markers linked with crown rust in Forage and Turf Breeding. New York: Springer; 2010. p. 331–40. Arojju et al. BMC Genetics (2018) 19:35 Page 10 of 10 45. Roderick HW, Thomas BJ. Infection of ryegrass by three rust fungi (Puccinia coronata, P graminis and P loliina) and some effects of temperature on the establishment of the disease and sporulation. Plant Pathol. 1997;46(5):751–61. 46. Roderick HW, Thorogood D, Adomako B. Temperature-dependent resistance to crown rust infection in perennial ryegrass, Lolium perenne. Plant Breed. 2000;119(1):93–5. 47. Schein RD, Rotem J. Temperature and humidity effects on uredospore viability. Mycologia. 1965;57(3):397–403. 48. Fè D, Pedersen MG, Jensen CS, Jensen J. Genetic and environmental variation in a commercial breeding program of perennial ryegrass. Crop Sci. 2015;55(2):631–40. 49. Muranty H, Troggio M, Sadok IB, Al Rifaï M, Auwerkerken A, Banchi E, Velasco R, Stevanato P, Van De Weg WE, Di Guardo M, et al. Accuracy and responses of genomic selection on key traits in apple breeding. Hortic Res. 2015;2:15060. 50. Wimmer V, Lehermeier C, Albrecht T, Auinger H-J, Wang Y, Schön CC. Genome-wide prediction of traits with different genetic architecture through efficient variable selection. Genetics. 2013;195(2):573–87. 51. Isidro J, Jannink J-L, Akdemir D, Poland J, Heslot N, Sorrells ME. Training set optimization under population structure in genomic selection. Theor Appl Genet. 2015;128(1):145–58. 52. Lorenz AJ, Smith KP, Jannink JL. Potential and optimization of genomic selection for fusarium head blight resistance in six-row barley. Crop Sci. 2012;52(4):1609–21. 53. Riedelsheimer C, Endelman JB, Stange M, Sorrells ME, Jannink J-L, Melchinger AE. Genomic predictability of interconnected biparental maize populations. Genetics. 2013;194(2):493–503. 54. Tayeh N, Klein A, Le Paslier MC, Jacquin F, Houtin H, Rond C, Chabert-Martinello M, Magnin-Robert JB, Marget P, Aubert G, et al. Genomic prediction in pea: effect of marker density and training population size and composition on prediction accuracy. Front Plant Sci. 2015;6:941. 55. Habier D, Fernando RL, Dekkers JCM. The impact of genetic relationship information on genome-assisted breeding values. Genetics. 2007;177(4):2389–97. 56. Zhang Z, Ding X, Liu J, Zhang Q, de Koning D-J. Accuracy of genomic prediction using low-density marker panels. J Dairy Sci. 2011;94(7): 3642–50. 57. Habier D, Fernando RL, Dekkers JCM. Genomic selection using low-density marker panels. Genetics. 2009;182(1):343–53. 58. Byrne SL, Nagy I, Pfeifer M, Armstead I, Swain S, Studer B, Mayer K, Campbell JD, Czaban A, Hentrup S, et al. A synteny-based draft genome sequence of the forage grass lolium perenne. Plant J. 2015;84(4):816–26. 59. Pfeifer M, Martis M, Asp T, Mayer KF, Lübberstedt T, Byrne S, Frei U, Studer B. The perennial ryegrass genomezipper: targeted use of genome resources for comparative grass genomics. Plant Physiol. 2013;161(2): 571–82. 60. Postel S, Küfner I, Beuter C, Mazzotta S, Schwedt A, Borlotti A, Halter T, Kemmerling B, Nürnberger T. The multifunctional leucine-rich repeat receptor kinase bak1 is implicated in arabidopsis development and immunity. Eur J Cell Biol. 2010;89(2):169–74. 61. Campbell NR, Harmon SA, Narum SR. Genotyping-in-thousands by sequencing (GT-seq): A cost effective SNP genotyping method based on custom amplicon sequencing. Mol Ecol Resour. 2015;15(4):855–67. 62. Proost S, Van Bel M, Vaneechoutte D, Van de Peer Y, Inzé D, Mueller-Roeber B, Vandepoele K. Plaza 30: an access point for plant comparative genomics. Nucleic Acids Res. 2015;43(D1):974–81.

Journal

BMC GeneticsSpringer Journals

Published: May 29, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off