Theoretical and Applied Genetics (2022) 135:3211–3222 https://doi.org/10.1007/s00122-022-04180-2 ORIGINAL ARTICLE Transposable element polymorphisms improve prediction of complex agronomic traits in rice Ioanna‑Theoni Vourlaki1,2 · Raúl Castanera2 · Sebastián E. Ramos‑Onsins2 · Josep M. Casacuberta2 · Miguel Pérez‑Enciso1,2,3 Received: 20 January 2022 / Accepted: 12 July 2022 / Published online: 5 August 2022 © The Author(s) 2022 Abstract Key message Transposon insertion polymorphisms can improve prediction of complex agronomic traits in rice com- pared to using SNPs only, especially when accessions to be predicted are less related to the training set. Abstract Transposon insertion polymorphisms (TIPs) are significant sources of genetic variation. Previous work has shown that TIPs can improve detection of causative loci on agronomic traits in rice. Here, we quantify the fraction of variance explained by single nucleotide polymorphisms (SNPs) compared to TIPs, and we explore whether TIPs can improve predic- tion of traits when compared to using only SNPs. We used eleven traits of agronomic relevance from by five different rice population groups (Aus, Indica, Aromatic, Japonica, and Admixed), 738 accessions in total. We assess prediction by applying data split validation in two scenarios. In the within-population scenario, we predicted performance of improved Indica varie- ties using the rest of Indica accessions. In the across population scenario, we predicted all Aromatic and Admixed accessions using the rest of populations. In each scenario, Bayes C and a Bayesian reproducible kernel Hilbert space regression were compared. We find that TIPs can explain an important fraction of total genetic variance and that they also improve genomic prediction. In the across population prediction scenario, TIPs outperformed SNPs in nine out of the eleven traits analyzed. In some traits like leaf senescence or grain width, using TIPs increased predictive correlation by 30–50%. Our results evi- dence, for the first time, that TIPs genotyping can improve prediction on complex agronomic traits in rice, especially when accessions to be predicted are less related to training accessions. Introduction of climate change and limited cultivable land. Current pace of rice genetic improvement may be too slow to meet these More than half of the world population consumes rice demands (Rosegrant and Cline 2003; Zhao et al. 2018). (Oryza sativa) in their daily diet. To secure nutritional Genomic selection can be a useful tool to accelerate genetic requirements of a growing human population, the improve- progress (Meuwissen et al. 2001). Numerous studies in rice ment of grain yield, both in quantity and in nutritional qual- and in other plant species (Jighly et al. 2019; Tessema et al. ity, is imperative. This is a significant challenge in the face 2020; Xu et al. 2020; Krishnappa et al. 2021) have already shown that genomic prediction (GP) can increase breeding speed. GP is particularly effective when traits are controlled Communicated by Hiroyoshi Iwata. by numerous loci which are difficult to map individually, * Ioanna‑Theoni Vourlaki such as yield and other traits of agronomic interest. For a ioanna.vourlaki@cragenomica.es recent review in rice, see Xu et al. (2021). * Miguel Pérez‑Enciso Conceptually, genomic prediction (GP) is a “large p, miguel.perez@uab.es small n” scenario where the number of variables p (molecu- lar markers) is typically far larger than the number of obser- 1 Universitat Autònoma de Barcelona, Department of Animal vations n. In this setting, either variables must be selected or Production, 08193 Bellaterra, Barcelona, Spain restrictions on the solutions must be imposed, or a combi- 2 Centre for Research in Agricultural Genomics nation of both. Methods such as LASSO (Tibshirani 2011) CSIC-IRTA-UAB-UB, 08193 Bellaterra, Barcelona, Spain or Bayes C (Habier et al. 2011) are examples of the first 3 Catalan Institute for Research and Advanced studies, ICREA, choice, whereas ridge regression or GBLUP (VanRaden 08010 Barcelona, Spain 13 Vol.:(0123456789) 3212 Theoretical and Applied Genetics (2022) 135:3211–3222 2008) involve restrictions on the square of solutions (L2 improve prediction accuracy. For this purpose, we used the norm). Numerous metrics exist for measuring predictive TIP genotypes from Castanera et al. (2021) and the phe- ability. Among others, it can be measured as the correlation notype database hosted in IRRI (Jackson 1997; Mansueto between predicted and observed phenotypes by splitting the et al. 2017). Note that a better model fit, as observed in data in training and test sets. Prediction accuracy is affected GWAS, does not necessarily imply a more accurate predic- by different factors such as the size of the training data, her- tion and thus the question posed here is pertinent. Further, itability, similarity between training and testing populations, any improvement in prediction albeit small can translate into or choice of marker sets (Goddard and Hayes 2007; Robert- large genetic gains when accumulated through generations. sen et al. 2019; Xu et al. 2021). In general, there is no consensus on which GP method is best. A recent review by Reinoso-Peláez et al. (2022) Materials and methods points at Reproducible Kernel Hilbert Space (RKHS) as the best overall method in plants. But there is variability. For Rice accessions and traits instance, Tehseen et al. (2021) compared GBLUP, Ridge Regression (RR), LASSO, Elastic Net (EN), Bayesian Ridge The 738 accessions used in this study (Supplementary Regression (BRR), Bayesian alphabet (A, B, C, …), RKHS Table 1) are from the collection conserved at IRRI used for for different traits, observing that no single method outper- the 3000-rice genome project (Jackson 1997; Li et al. 2014) formed the rest for all traits. Kaler et al. (2022) conducted a and were chosen because they were sequenced at least at comparative study among 11 different methods for two traits 15 × depth. The 738 accessions retained pertain to all main in soybean, rice, and maize, reporting better predictive abili- rice population groups: Aus/Boro (AUS, N = 75), Indica ties using Bayes B. Xu et al. (2018) found that GBLUP and (IND, N = 451), Japonica (JAP, N = 166), Aromatic (ARO, LASSO performed best in hybrid breeding. Other authors N = 17). The accessions that cannot be assigned to a specific have suggested integrating genomic prediction with crop rice group are categorized as Admixed (ADM, N = 29). We growth models to evaluate the efficiency of phenotypic strat- used the SNP-based group assignment from Sun et al. (2017) egies and the impact of the different yield components on to identify the different subsets of this study. the prediction accuracy (Bustos-Korts et al. 2019; Cooper Out of the 56 traits originally available at IRRI SNP- et al. 2016). Selecting SNPs based on genome-wide associa- Seek database (https://​snp-​seek.​irri.​org/), we chose the 11 tion studies (GWAS) has also been proposed, e.g., Spindel traits for which data were available in the 738 accessions et al. (2016). selected. Some discrete traits were binned to balance the Irrespective of the algorithm chosen, single nucleotide number of observations per class and time to flowering was polymorphisms (SNPs) are the main class of markers used log-transformed. Supplementary Table 2 shows basic sta- so far in GP due to their genome-wide abundance and geno- tistics and transformations applied. Principal component typing automatization. SNPs are not, however, the only analysis (PCA) for the 11 phenotypes was obtained with the source of phenotypic variability in the genome. In the last “prcomp” function available in R.4.1.0 (Team 2021) envi- few years, data have accumulated on the importance of pres- ronment. For plotting loading variables of PCA, package ence–absence variation and structural variation as a source “factorextra” (Kassambara and Mundt 2020) and packages of phenotypic variability in plants, including in rice (e.g., “ggrepel” (Slowikowski 2020) and “ggbiplot” (Vu 2011) for Fuentes et al. 2019). Transposon insertion polymorphisms the biplot were used. (TIPs) can account for a major fraction of intraspecific structural variation, as it has been recently found in maize Markers (Haberer et al. 2020). In fact, transposable elements are con- sidered as one of the main drivers for plant genome vari- A binary ped file format with the Core SNP dataset for all ability, impacting on genome coding capacity and regulation chromosomes was downloaded from the SNP-Seek database. in numerous ways (Lisch 2013). However, until the recent The original dataset consisted of 404,388 bi-allelic SNPs development and evaluation of reliable methods for calling from 3,034 rice accessions, including the 738 accessions TIPs from short-read resequencing data (Vendrell-Mir et al. selected. Markers with minor allele frequency ≤ 0.01 and 2019), it was not possible to use TIPs for GWAS approaches. missing rate > 1% were filtered out using plink2 (Purcell Importantly, recent studies in rice and in tomato have et al. 2007; Chang et al. 2015). Missing genotypes were shown that the use of TIPs as genetic information can imputed using Beagle 5.2 with default parameters (Brown- result in an increase of association signals as compared ing et al. 2018). The final dataset consisted of 228,871 SNPs, to SNPs in GWAS (Carpentier et al. 2019; Akakpo et al. which were used for the analyses reported here. Of those, 2020; Domínguez et al. 2020; Castanera et al. 2021). These 50,485 SNPs were in gene regions. results prompt us to investigate whether transposons can also 13 Theoretical and Applied Genetics (2022) 135:3211–3222 3213 Transposable elements (TEs) are divided in two main implemented in BGLR package (Pérez and de Los Campos classes “copy and paste” (Class I TEs) or “cut and paste” 2014) using default priors to estimate 𝜎S2 , 𝜎M 2 , and 𝜎R2. Class II TEs. In rice, the most abundant Class I elements are RLX (LTR retrotransposons) and RIX (Non-LTR retrotrans- Genomic prediction posons), whereas DTX (DNA TEs with terminal inverted repeats) and MITEs (Miniature Inverted-repeat Transposable Plant breeding is primarily based on trials of new crosses, Elements) are the most prevalent (Mao et al. 2000). Here which can be lengthy and costly. The speed of development we used markers from both classes, accounting for 94% of for new improved varieties depends largely on accuracy of the TIPs described in Castanera et al. (2021). Class I TIPs prediction for new genotypes. We evaluated two distinct val- were represented by 21,571 RLX and RIX markers. Class idation scenarios that cover two important issues: prediction II consisted of 52,120 MITE and DTX markers. In contrast of performance within population (rice group in this case) to SNPs, TIPs can only be genotyped as presence/absence, and prediction of individuals from different groups. recoded consequently as 0/1, and defined as genomic win- In the first scenario, we measured accuracy when predict- dows with an average size of 1.2 kb. TIP windows were ing performance of improved Indica varieties (N = 48) using taken from Castanera et al. (2021) and are based on the inter- the rest of accessions, including non-improved Indica acces- section of the individual TE insertion regions predicted for sions. Accessions from IRRI core collection are classified as each accession with genome-wide windows of a fixed size “improved,” “breeding and inbred lines,” and “traditional” (1 kb, merging adjacent windows). These TIPs were further varieties. We used this passport information to identify this classified as genic or intergenic by intersecting the windows subset of improved varieties. “Improved” Indica varieties with MSU7 non-TE gene annotation (Kawahara et al. 2013). correspond to most modern and commercial lines avail- A TIP was considered genic if the window overlapped at able at IRRI collection. In this scenario, performance to be least 1 bp with the gene feature. There were 17,034 genic predicted is from highly related accessions to those in the MITE/DTX and 5,024 genic RLX/RIX TIPS. The remaining training set. TIPs were considered intergenic. In the second scenario, we predicted performance of MITEs amplify by bursts from individual elements creat- all Admixed (ADM, N = 29) and Aromatic (ARO, N = 17) ing highly homogeneous families, as previously reported in accessions using the rest of groups. In this case, performance Arabidopsis (Santiago et al. 2002) and rice (Lu et al. 2017). to be predicted is from accessions that may not be too related Different bursts of amplification at different evolutionary to accessions in the training set, and we expect prediction times may have different prediction potential for particular to be worse than in the former scenario. For instance, the phenotypes. In an attempt to study the potential predictive ADM group is a small, highly heterogeneous collection of capacity of individual families, we created individual TIP accessions. genotype matrices for each of the 18 largest MITE fami- The rationale for the first scenario is that new selected lies described in Castanera et al. (2021) (Supplementary accessions can be crosses within the same population, and Table 3). Each of these matrices included only TIPs origi- the breeder can be interested in designing new better per- nated from a single transposon, in this case MITE, family. forming crosses out of traditional varieties. The second sce- nario is more challenging, since we do not use any sample of the population to be predicted. These two scenarios, within Genetic variance inference and across populations, resemble main challenges faced in a breeding program. Note there are infinite designs for assess- We fitted the following linear model in order to estimate ing predictive accuracy. For instance, we did not study pre- genetic variance components explained by each marker set: diction in Japonica because we preferred to focus on a larger number of traits, since genetic architecture is a main factor y = 𝜇 + Zu1 + Zu2 + Zu3 + e (1a) influencing predictive performance (Daetwyler et al. 2010). where μ is the general mean, y is the phenotype vector of Ample literature shows that no single method performs size n, the number of accessions, Z is an identity incidence best for all traits and scenarios. Here, we compared two alter- matrix, u1, u2, u3 are random effects representing each of the native modeling strategies: Bayesian RKHS as described marker groups, and e is the residual. We assume u1 ~ N(0, above, and Bayes C. RKHS with the kernel employed here KS 𝜎S2 ), u2 ~ N(0, KM 𝜎M2 ), and u3 ~ N(0, KR 𝜎R2 ), where KS, is equivalent to ridge regression and GBLUP, whereas Bayes KM, KR are genomic relationship matrices obtained from C is a variable selection method. The two methods were SNPs, MITE/DTX, and RLX/RIX markers, respectively. applied to both predictive scenarios. For RKHS, we com- These matrices were calculated using AGHMatrix (Amadeu pared predictive performance using all markers (model 1a et al. 2016). Model 1a was fitted with a Bayesian Reproduc- above) with submodels ible Kernel Hilbert Space (RKHS, Herbrich et al. 1999) as 13 3214 Theoretical and Applied Genetics (2022) 135:3211–3222 be interpreted as the number of prior counts (prior ‘suc- y = 𝜇 + Zu1 + e, (1b) cesses’ plus prior ‘failures’).” Here we chose p0 = 5 and π0 = 0.01. y = 𝜇 + Zu2 + e, (1c) In a subset of cases, we evaluated whether using only and genic SNPs improved prediction compared to using all avail- able markers. Similarly, we conjectured that not all transpos- y = 𝜇 + Zu3 + e, (1d) able elements are equally likely to cause phenotypic changes. We analyzed predictive performance of models containing i.e., when using only SNPs (model 1b), only MITE/ TIPs from each of the largest 18 MITE families present in DTX (model 1c), or only RLX/RIX (model 1d) markers. the rice genome (Supplementary Table 3). To avoid repeti- For Bayes C, the complete model was: tive, lengthy results we make the additional analysis using y = 𝜇 + XS 𝛽1 + XM 𝛽2 + XR 𝛽3 + e, (2a) two agronomic traits of high importance on rice breeding, time to flowering, and grain length. An earlier or later grow- where XS, XM, and XR are the standardized genotypic ing can determine seed production. Grain size-related traits values of each marker class; β 1, β 2, and β 3 are the cor- such as grain length/width are important breeding targets responding vectors of effects for SNPs, MITE/DTX, and since they affect the quality of the crop yield. These two RLX/RIX markers, respectively. As with RKHS, three traits may also represent alternative genetic architecture partial models were also evaluated: (Begum et al. 2015; Xu et al. 2015; Chen et al. 2021). Using either RKHS or Bayes C, phenotypes to be pre- y = 𝜇 + XS 𝜷 1 + e, (2b) dicted were removed from the dataset, the model fitted using the remaining phenotypes, and the correlation between pre- y = 𝜇 + XM 𝜷 2 + e, (2c) dicted and observed phenotypes computed as a measure of and predictive accuracy. From a practical point of view, it is important to assess whether predictions using TIPs or all y = 𝜇 + XR 𝜷 3 + e. (2d) markers are better than the state-of-the-art method, i.e., with SNPs only. To assess variability of results, we generated In Bayes C, a probability π of presence/absence of a 10,000 bootstrap sampling with replacement from the cor- SNP in the model is sampled from π ∼ Beta(p0, π0). Fol- responding pairs of phenotypes observed and predicted with lowing Pérez and de Los Campos (2014, see their Tables 1 each method and marker set. We then computed the cor- and S1), “the beta prior is parameterized in a way that the relation observed–predicted samples within each bootstrap expected value by E(π) = π 0; on the other hand, p 0 can sample and we counted how many times correlation using SNPs only was lower than with each alternative strategy. Table 1  Means of posterior Traits All accessions (N = 738) Indica accessions (N = 451) distributions of genetic variances explained by each 𝜎S2 2 𝜎M 𝜎R2 𝜎S2 2 𝜎M 𝜎R2 marker set Culm diameter 0.16 0.17* 0.16 0.13 0.17* 0.15 Culm strength 0.10 0.25* 0.16 0.11 0.19* 0.14 Flag leaf angle 0.22* 0.14 0.15 0.24* 0.14 0.14 Grain length 0.48* 0.11 0.11 0.41* 0.11 0.13 Grain width 0.49* 0.11 0.12 0.42* 0.11 0.14 Leaf length 0.26* 0.16 0.19 0.22* 0.16 0.19 Leaf senescence 0.12 0.25* 0.18 0.14 0.21* 0.16 Grain weight 0.40* 0.11 0.13 0.31* 0.12 0.13 Salt injury 0.10 0.11 0.12* 0.09 0.11* 0.11* Time to flowering 0.45* 0.12 0.13 0.39* 0.13 0.13 Pan. threshability 0.11 0.13* 0.10 0.11 0.15* 0.11 𝜎S2: genetic variance explained by SNPs 2 𝜎M : genetic variance explained by DNA transposon markers (MITE/DTX) 𝜎R2 : genetic variance explained by retrotransposons (RLX/RIX) Traits are scaled such that phenotypic variances are 1 *Best strategy 13 Theoretical and Applied Genetics (2022) 135:3211–3222 3215 Phenotypic measurements and variables were centered and “heritability” because a proper interpretation assumes pan- scaled to mean 0 and variance 1. BGLR was run for 100,000 mixia, a condition not fulfilled here. Having these caution- iterations using default priors for RKHS. This number of ary remarks in mind, Table 1 shows that transposons can iterations seemed enough to attain convergence (Supplemen- explain a sizeable fraction of genetic variance, which was tary Fig. 1). larger than that explained by SNPs in five out of 11 traits. In seven traits, 𝜎S2 was smaller than the sum of 𝜎M 2 and 𝜎R2 . Results are presented when all accessions were analyzed Results together and when using only data from Indica, the largest group (N = 451). Note model (1a) assumes constant genetic Descriptive analysis variances across accessions, i.e., 𝜎S2, 𝜎M 2 , and 𝜎R2 are the same in all rice groups. This is not necessarily the case. Neverthe- Figure 1a shows the loadings, i.e., the projections of vari- less, variances were similar within Indica and across popula- ables into the lower-dimensional space, of each trait to the tion groups. principal components. In the figure, the length of the arrow is proportional to trait contribution and the angle between Genomic prediction arrows, to their correlation. An analysis in two principal components shows that the first component depends on We assess prediction in two validation scenarios that rep- grain width and grain weight, whereas culm diameter, time resent some of the main challenges in breeding, prediction to flowering, and leaf length are the main contributors to within and across populations (see “Materials and meth- the second component. The rest of traits contribute more ods”). In the first one, Indica improved varieties were pre- modestly to total phenotypic variation. A sample projection dicted using the rest of accessions, including traditional (Fig. 1b) shows graphically how accessions differed in the Indica varieties. In this scenario, using TIPs increased pre- traits studied. Supplementary Fig. 2 shows the differences in diction accuracy compared to using SNPs in six traits: culm trait distributions across accessions. In general, populations diameter, grain length, leaf length, leaf senescence, grain differed for most traits although to varying extent. Figure 1b weight, and time to flowering (Fig. 2). indicates, e.g., that Japonica accessions tend to have higher In the second scenario, phenotypes of all ADM and ARO grain weights and widths, as they are projected in the lower accessions were predicted given the rest of the accessions. part of the figure, and as shown in Supplementary Fig. 2. TIPs were especially beneficial in this case: TIPs improved prediction upon using only SNPs in nine out of the 11 traits Genetic variance estimates analyzed (Fig. 3). In some traits, such as grain width or leaf senescence, improvement in correlation using TIPs was The genetic variance explained by each marker set meas- remarkable, over 30%. In other traits, such as time to flow- ures its relative importance in determining the observed ering, improvement was marginal. For some traits, notably phenotypes. Here we prefer not to use the classical term Fig. 1  a PC loadings of each trait for the two first standardized principal components. b Plot showing the accessions projected. The first (x-axis) and second (y-axis) PCs explained 19% and 15.8% of variance, respectively 13 3216 Theoretical and Applied Genetics (2022) 135:3211–3222 Fig. 2  Correlation between observed and predicted phenotypes of utilized: green, SNPs; magenta, MITE/DTX; blue, RLX/RIX; brown, Indica improved varieties. In each plot, the first four columns repre- all markers. The asterisk shows the best option for each trait. (Color sent the correlation values using Bayes C, while the last four values figure online) correspond to RKHS method. Colors represent marker information grain weight and panicle threshability, prediction across both contribute largely to total phenotypic variation (Fig. 1). populations was successful neither with SNPs nor with TIPs. This suggests that genomic prediction combined with trans- We computed the bootstrap probability that using TIPs, posable elements can be an effective tool for overall rice or all markers resulted in better predictions than using only genetic improvement as it would enhance genetic progress SNPs (see “Materials and methods”). Results are given in in important agronomic traits. Supplementary Tables 4 and 5 for the within and across Note that using all markers is not necessarily the best population scenarios, respectively. Even if gains in accuracy option for predictive purposes: it only outperformed the shown in Figs. 2, 3 may seem small in some cases, results rest of approaches in three out of the 44 (= 11 traits × 2 are consistent. For instance, increase in correlation for leaf methods × 2 predictive scenarios) analyses. This indi- length is ~ 15% when using MITE/DTX compared to SNPs cates that adding additional markers may contribute to in the within population scenario, a somewhat modest figure. overfitting and reduce model performance in prediction. But this result is confirmed in 80% of the bootstrap samples. Overfitting is a well-known phenomenon in the machine In contrast, SNPs are far better than TIPs for culm strength learning literature when the model is not properly regular- and this is also confirmed in bootstrap samples (Supplemen- ized. This has been clearly observed with simulated data tary Table 4, Fig. 2). in a genomic prediction scenario (e.g., Pérez-Enciso et al. On average, prediction across populations was less accu- 2015). rate than within Indica in seven out of 11 traits and irrespec- Next, we wished to study how the different genetic archi- tive of marker set used (Figs. 2, 3). Importantly, gain using tectures influence the statistical behavior of the three sets TIPs was larger in this scenario than in the within popula- of markers. Bayes C is a variable selection method and so tion scenario. Time to flowering and grain width were the we reasoned that the number of markers entering the model traits for which prediction was most accurate. Nevertheless, and their effects would differ between traits. Broadly, esti- prediction across populations for grain width was far less mates of marker effects were quite similar across traits (for precise than within Indica. It is interesting to note that grain the same type of marker) as can be seen in Supplementary width and time to flowering are basically uncorrelated, but Fig. 3. The only exception was grain width and grain length, 13 Theoretical and Applied Genetics (2022) 135:3211–3222 3217 Fig. 3  Correlation between observed and predicted phenotypes across RKHS method. Colors represent marker information utilized: green, accessions. All ADM and ADM accessions were predicted using the SNPs; magenta, MITE/DTX; blue, RLX/RIX; brown, all markers. rest of groups. In each plot, the first four columns represent the corre- The asterisk shows the best option for each trait. (Color figure online) lation values using Bayes C, while the last four values correspond to where we observed much larger estimated effects for MITE/ Using a subset of all markers available can improve DTX and SNPs, respectively, in agreement with results in prediction. For instance, the accuracy of a model which Fig. 3. In turn, there were larger differences between the contains only the causative SNPs can approach one (Pérez- probabilities (d) of entering the model for each marker type Enciso et al. 2015). The problem, of course, is that causative (Supplementary Fig. 4). This occurred despite setting equal mutations cannot be identified in most cases. Several indi- priors for all types of markers (p = 0.01). This was not due rect approaches have been suggested instead. For instance, only to the priors or different number of TIPs compared to Spindel et al. (2016) proposed to perform prediction using SNPs, because the pattern differed between traits. the most associated markers, e.g., selected via a GWAS Table 2  Predictive accuracy when using all or only gene-based markers Prediction scenario Traita Markers Bayes C RKHS SNPS MITE/DTX RLX/RIX ALL SNPS MITE/DTX RLX/RIX ALL Across GL Genic 0.57 0.51 0.41 0.67* 0.47 0.54 0.26 0.48 All 0.65 0.42 0.42 0.66 0.45 0.50 0.35 0.48 TF Genic 0.71 0.58 0.50 0.68 0.71 0.65 0.69 0.75* All 0.71 0.66 0.56 0.71 0.68 0.72 0.67 0.73 Within GL Genic 0.56 0.55 0.43 0.57 0.56 0.38 0.36 0.55 All 0.61 0.69* 0.43 0.63 0.59 0.50 0.45 0.59 TF Genic 0.59 0.55 0.51 0.62 0.57 0.61 0.56 0.57 All 0.59 0.63* 0.62 0.60 0.59 0.65 0.59 0.60 *Best strategy a GL: grain length; TF: time to flowering 13 3218 Theoretical and Applied Genetics (2022) 135:3211–3222 P-values. We did not evaluate this strategy here, although Discussion we did consider two alternative approaches for preselect- ing markers. In a first attempt, we examined whether using We have shown, for the first time to our knowledge, that only gene-based markers improved prediction performance. transposable element polymorphisms can improve predictive To avoid multiplying analyses, we selected grain length and accuracy for important agronomic traits in rice. The impact time to flowering. As can be seen in Table 2, gene-based of using TIPs varied; here we found that they improved markers outperform all markers in across population but predictive performance in ~ 60% of the traits and scenarios minimally. The opposite was observed in the within popu- considered. Table 3 presents a summary. The increase in lation scenario. accuracy also varied. Although the added benefit of using We also studied performance of TIPs pertaining to each TIPs was sometimes modest, TIPs improved correlation by of the 18 largest MITE families (Supplementary Table 3, more than 30% in traits like grain width or leaf senescence. see “Materials and methods”). Again, for brevity, we con- All traits analyzed here have an economic impact in rice sidered only prediction across accessions in grain length and production. Unfortunately though, grain yield phenotypic time to flowering using RKHS (Fig. 4). The most relevant data are not available for the 3 k rice panel, and how grain conclusion is that predictive performance can vary largely yield is affected by TIPs remains to be studied. This trait according to MITE family and that using SNPs on top of is largely affected by genotype × environment interaction, MITEs may not improve prediction. Prediction of time to and so the relevance of TIPs may be harder to characterize. flowering improved using MITE family MH63fam47_235 Among the traits studied, time to flowering is particularly (MITE-adh type B-like superfamily) TIPs compared to using important (Wang and Li 2019). Rice plants needs approxi- the full MITE/DTX set (Figs. 3, 4). Although it is tempting mately, 3–6 months to grow, meaning that earlier or later to conclude that a specific MITE family is enriched in genes growing can strongly affect the yield. Productivity is also affecting a given trait, one should be careful as disequilib- determined by morphological trait such as grain weight rium can extend over long genome regions (Mather et al. (Chen et al. 2021). Grain weight in turn correlates with 2007; Nachimuthu et al. 2015). grain width, Fig. 1 (Li et al. 2021). Most of these traits are polygenic. Some traits like time to flowering, grain weight, grain width, and grain length seem controlled by large effect quantitative trait loci (Begum et al. 2015; Xu et al. 2015; Chen et al. 2021). For some traits, e.g., grain width, GP was Fig. 4  Predictive accuracy across populations using TIPS from each of 18 recognized MITE families. Each column corresponds to accuracy with one MITE family. Model included only MITEs or MITEs and all SNPs. The asterisk shows the best option 13 Theoretical and Applied Genetics (2022) 135:3211–3222 3219 Table 3  Maximum predictive Trait Scenario accuracy and corresponding marker set Indica improved varieties ARO/ADM accessions Culm diameter 0.40 (MITE/DTX) 0.26 (MITE/DTX) Culm strength 0.28 (SNPs) 0.16 (RLX/RIX) Flag leaf angle 0.45 (SNPs) 0.28 (RLX/RIX) Grain length 0.69 (MITE/DTX) 0.66 (ALL) Grain width 0.83 (SNPs, ALL) 0.64 (MITE/DTX) Leaf length 0.41 (MITE/DTX) 0.52 (RLX/RIX) Leaf senescence 0.47 (RLX/RIX) 0.54 (MITE/DTX, RLX/RIX) Grain weight 0.30 (MITE/DTX) 0.14 (MITE/DTX) Salt injury 0.28 (SNPs) 0.49 (SNPs, MITE/DTX) Time to flowering 0.65 (MITE/DTX) 0.73 (ALL) Panicle threshability 0.29 (SNPs) 0.24 (SNPs) quite accurate and we confirm that GP can largely enhance for some agronomic traits. Importantly, we found TIPs are rice genetic progress, in agreement with previous results (Xu especially helpful when prediction was across populations. et al. 2021). For other traits, e.g., leaf length, GP accuracy These issues merit further research. was lower, although it is interesting to note that bootstrap The main families of class I in rice are LTR-retrotranspo- sampling suggests that results are repeatable (Supplemen- sons (RLX) and LINEs (RIX), whereas DNA transposons tary Tables 4, 5). Since plant breeding builds on cumulative (DTX) and MITEs are the main components of rice class II progress over generations, even a small advantage can be TEs (Matsumoto et al. 2005). There are important structural highly relevant in the medium to long term. and mechanistic differences between class I, or retrotrans- The reasons behind the high capacity of TIPs to predict posons, and class II, or DNA transposons. Although both phenotypes, which in some cases is far better than SNPs, RLX and MITEs target genic regions for integration, their could be manifold. Transposable element insertions can have dynamics is very different. While RLXs have a high turnover stronger effects than SNPs as some transposon types tend to and RLX TIPs are usually present at a very low frequency localize near genes. Therefore, TIPs could be in some cases in the population, MITEs are maintained in the genome for causative mutations linked to a specific trait. Indeed, trans- longer evolutionary periods (Castanera et al. 2021). This posable element insertions are known to have played a major suggests that although both types of TEs can be associ- role in plant genome evolution both in the wild and under ated with traits in rice (Akakpo et al. 2020; Castanera et al. breeding settings, and examples of TIP causative mutations 2021), their capacity to predict phenotypes may differ. Cer- for many agricultural important traits have been reported tainly, our results show that MITE/DTX are more relevant (Lisch 2013; Dubin et al. 2018). In some cases, the TIPs than RLX/RIX for improving prediction (Table 3, Figs. 2, linked to the trait may be recent insertions and may not be 3). It is finally interesting to note that a single MITE family in high LD with surrounding SNPs. This is what was shown of ~ 3 k TIPs can predict equally well a phenotype as well as in recent GWAS analyses performed with TIPs and SNPs in 200 k SNPs (Fig. 4). In contrast, we did not find a consistent tomato and rice, where TIPs revealed associations with traits or large improvement in prediction when using only gene that are not detected with SNPs (Domínguez et al. 2020; markers as compared to using all available polymorphisms, Akakpo et al. 2020; Castanera et al. 2021). In contrast to as reported also in humans (Visscher et al. 2021). SNP mutation rate, transposon activity is not constant over Some technical considerations should be borne in mind time, with bursts of transposition associated with stress situ- regarding our analyses. Ordinal traits (Supplementary ations or environmental stimuli (Dubin et al. 2018). There- Fig. 2) were treated as continuous. It has been known for fore, it can be hypothesized that the adaptation of a crop decades that a threshold (logistic model) is theoretically a to a new environment, say as part of the breeding process, better choice for binary traits than standard linear models could be a period particularly prone to transposition activ- (Gianola and Foulley 1983). The logistic model is a class ity (Baduel and Quadrana 2021). On the other hand, while of the so-called generalized linear models, where the non- SNPs accumulate relatively homogeneously throughout the linear relationship between parameters and observations genome, some TEs target gene-rich regions for integration, becomes linear after applying a transformation, e.g., logit for particularly RLXs and MITEs in rice (Castanera et al. 2021). binary traits. Despite their theoretical appeal, these models Therefore, the potential for TEs to produce causal mutations are more difficult to run than linear counterparts and may and TIP associations with traits could be particularly high converge poorly. Empirical evidence generally shows small 13 3220 Theoretical and Applied Genetics (2022) 135:3211–3222 differences only (Matos et al. 1997; Olesen et al. 1994). Supplementary Information The online version contains supplemen- Here, we observed (Supplementary Table 4) that a threshold tary material available at https://d​ oi.o​ rg/1​ 0.1​ 007/s​ 00122-0​ 22-0​ 4180-2. model may have a small advantage over linear ones but not Author contribution statement MPE, JMC, and SERO designed and always. A second issue is the metrics to assess prediction. supervised work; ITV, RC, and MPE wrote code; ITV and RC ana- Here we chose correlation as it has a direct interpretation in lyzed data; MPE and ITV wrote the manuscript, with input from rest terms of response to selection (Falconer and Mackay 1996) of authors. and has been widely used, but numerous other metrics exist. For instance, mean square error (RMSE) of prediction is Funding Open Access Funding provided by Universitat Autonoma de Barcelona. The project was funded by Ministry of Science and also widely used. We computed RMSE (Supplementary Innovation-State Research Agency (AEI, Spain, https://​doi.​org/​10.​ Table 7, 8) and we found concordant results regarding the 13039/​50110​00110​33) grant numbers PID2019-106374RB-I00 to best marker set in 9 (within scenario) or 10 traits (across JMC, PID2020-119255 GB-I00 to SERO and PID2019-108829RB- scenario) out of the 11 traits studied. These issues do not I00 to MPE. ITV is supported by a predoctoral fellowship funded by MCIN/AEI/https://​doi.​org/​10.​13039/​50110​00110​33 through the Grant question our main, and most important conclusion regarding BES-2017–081139 and by “ESF Investing in your future.” RC holds a that TIPs can improve genomic prediction. Juan de la Cierva Incorporación Postdoctoral fellowship funded by the A prerequisite for the inclusion of TIPs in practical Spanish Ministry of Science and Innovation-State Research Agency. breeding programs is to automatize their genotyping. TIP This work was also supported by grant CEX2019-000902-S funded by MCIN/AEI/https://​doi.​org/​10.​13039/​50110​00110​33 and by the genotyping should primarily target high-frequency TIPs CERCA Programme/Generalitat de Catalunya (Spain). in order to be as informative as possible, as it is usually done for SNPs as well. The application of TIP-Chip (Whee- Availability of data and materials All data generated and software used lan et al. 2006) or transposon insertion profiling (TIP-seq, during this study are included in a Github site https://​github.​com/​ivour​ Steranka et al. 2019), and TE-sequence capture (Quadrana laki/​Trans​posab​le-​eleme​nt-​polym​orphi​sms-​impro​ve-​predi​ction-​of-​ compl​ex-​agron​omic-​traits-​in-​rice.​git. et al. 2021) to hundreds or thousands of varieties should be cheap, as the sequencing coverage needed per sample is very Declarations small. Finally, given the dropping costs of genome sequenc- ing, thousands of rice accessions are being re-sequenced Conflict of interest The authors declare that they have no conflict of and made public. TIPs could also be included in standard interest. genotyping arrays (Wheelan et al. 2006) as a complement to SNPs. Given that TIPs from a single MITE family can be as Open Access This article is licensed under a Creative Commons Attri- efficient as 200 k SNPs in some traits (Fig. 4), perhaps only bution 4.0 International License, which permits use, sharing, adapta- tion, distribution and reproduction in any medium or format, as long a small number of TIPs need to be included in the genotyp- as you give appropriate credit to the original author(s) and the source, ing protocol. provide a link to the Creative Commons licence, and indicate if changes In conclusion, we consistently observed that TIPs can were made. The images or other third party material in this article are increase predictive accuracy of agronomic traits in rice and included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in do explain a non-negligible fraction of phenotypic vari- the article's Creative Commons licence and your intended use is not ance. Notably, this improvement was larger when prediction permitted by statutory regulation or exceeds the permitted use, you will was across populations than within Indica. Using markers need to obtain permission directly from the copyright holder. To view a positioned within genes did not seem to matter too much, copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. although perhaps a more thorough analysis would be needed. In contrast, selecting TIPs from some transposon families did improve prediction. These are important results from a References practical point of view and warrant developments to automa- Akakpo R, Carpentier MC, Ie Hsing Y, Panaud O (2020) The impact tize TIP genotyping. From a biological point of view, new of transposable elements on the structure, evolution and function studies are needed to understand how TIPs affect complex of the rice genome. New Phytol 226:44–49 trait variation. Improving predictive accuracy from molecu- Amadeu RR, Cellon C, Olmstead JW et al (2016) AGHmatrix: R pack- lar data is an important task since even small gains add up age to construct relationship matrices for autotetraploid and dip- loid species: a blueberry example. Plant Genome. https://​doi.​org/​ over generations and can make a big long-term difference. 10.​3835/​plant​genom​e2016.​01.​0009 Assessing the importance of TIPs in other agronomic traits, Baduel P, Quadrana L (2021) Jumpstarting evolution: how transposi- such as grain yield across different environments, remains tion can facilitate adaptation to rapid environmental changes. Curr also to be studied. Once a plausible set of parameters linking Opin Plant Biol 61 Begum H, Spindel JE, Lalusin A et al (2015) Genome-wide association TIPs, SNPs, and yield are estimated from real data, simula- mapping for yield and other agronomic traits in an elite breeding tions can be used to optimize marker genotyping with SNPs population of tropical rice (Oryza sativa). PLoS ONE. https://d​ oi.​ and/or TIPs. org/​10.​1371/​journ​al.​pone.​01198​73 13 Theoretical and Applied Genetics (2022) 135:3211–3222 3221 Browning BL, Zhou Y, Browning SR (2018) A one-penny imputed Kawahara Y, de la Bastide M, Hamilton JP et al (2013) Improvement of genome from next-generation reference panels. Am J Hum Genet the oryza sativa nipponbare reference genome using next genera- 103:338–348. https://​doi.​org/​10.​1016/j.​ajhg.​2018.​07.​015 tion sequence and optical map data. Rice 6:3–10. https://​doi.​org/​ Bustos-Korts D, Boer MP, Malosetti M et al (2019) Combining crop 10.​1186/​1939-​8433-6-4 growth modeling and statistical genetic modeling to evaluate phe- Krishnappa G, Savadi S, Tyagi BS et al (2021) Integrated notyping strategies. Front Plant Sci. https://​doi.​org/​10.​3389/​fpls.​ genomic selection for rapid improvement of crops. Genomics 2019.​01491 113:1070–1086 Carpentier MC, Manfroi E, Wei FJ et al (2019) Retrotranspositional Li Z, Fu BY, Gao YM et al (2014) The 3,000 rice genomes project. landscape of Asian rice revealed by 3000 genomes. Nat Commun Gigascience 3:2047–2217. https://d​ oi.o​ rg/1​ 0.1​ 186/2​ 047-2​ 17X-3-7 10:24. https://​doi.​org/​10.​1038/​s41467-​018-​07974-5 Li G, Tang J, Zheng J, Chu C (2021) Exploration of rice yield poten- Castanera R, Vendrell-Mir P, Bardil A et al (2021) Amplification tial: Decoding agronomic and physiological traits. Crop Journal dynamics of miniature inverted-repeat transposable elements and 9:577–589 their impact on rice trait variability. Plant J 107:118–135. https://​ Lisch D (2013) How important are transposons for plant evolution? doi.​org/​10.​1111/​tpj.​15277 Nat Rev Genet 14:49–61 Chang CC, Chow CC, Tellier LCAM et al (2015) Second-generation Lu L, Chen J, Robb SMC et al (2017) Tracking the genome-wide out- PLINK: Rising to the challenge of larger and richer datasets. comes of a transposable element burst over decades of amplifica- Gigascience. https://​doi.​org/​10.​1186/​s13742-​015-​0047-8 tion. Proc Natl Acad Sci USA 114:E10550–E10559. https://​doi.​ Chen K, Łyskowski A, Jaremko Ł, Jaremko M (2021) Genetic and org/​10.​1073/​pnas.​17164​59114 molecular factors determining grain weight in rice. Front Plant Mansueto L, Fuentes RR, Borja FN et al (2017) Rice SNP-seek data- Sci. https://​doi.​org/​10.​3389/​fpls.​2021.​605799 base update: new SNPs, indels, and queries. Nucleic Acids Res Cooper M, Technow F, Messina C et al (2016) Use of crop growth 45:D1075–D1081. https://​doi.​org/​10.​1093/​nar/​gkw11​35 models with whole-genome prediction: application to a maize Mao L, Wood TC, Yu Y et al (2000) Rice Transposable elements: multienvironment trial. Crop Sci 56:2141–2156. https://​doi.​org/​ a survey of 73,000 sequence-tagged-connectors. Genome Res 10.​2135/​crops​ci2015.​08.​0512 10:982–990. https://​doi.​org/​10.​1101/​gr.​10.7.​982 Daetwyler HD, Pong-Wong R, Villanueva B, Woolliams JA (2010) Mather KA, Caicedo AL, Polato NR et al (2007) The extent of linkage The impact of genetic architecture on genome-wide evaluation disequilibrium in rice (Oryza sativa L.). Genetics 177:2223–2232. methods. Genetics 185:1021–1031. https://​doi.​org/​10.​1534/​genet​ https://​doi.​org/​10.​1534/​genet​ics.​107.​079616 ics.​110.​116855 Matos CAP, Thomas DL, Gianola D et al (1997) Genetic analysis of Domínguez M, Dugas E, Benchouaia M et al (2020) The impact discrete reproductive traits in sheep using linear and nonlinear of transposable elements on tomato diversity. Nat Commun models: II. Goodness of fit and predictive ability. J Anim Sci 11:1–11. https://​doi.​org/​10.​1038/​s41467-​020-​17874-2 75:88–94. https://​doi.​org/​10.​2527/​1997.​75188x Dubin MJ, Mittelsten Scheid O, Becker C (2018) Transposons: a Matsumoto T, Wu J, Kanamori H et al (2005) The map-based blessing curse. Curr Opin Plant Biol 42:23–29 sequence of the rice genome. Nature 436:793–800. https://​doi.​ Falconer DS, Mackay TFC (1996) Introduction to quantitative genet- org/​10.​1038/​natur​e03895 ics, 4th edn. Addison Wesley Longman, Harlow Meuwissen THE, Hayes BJ, Goddard ME (2001) Prediction of total Fuentes RR, Chebotarov D, Duitama J et al (2019) Structural vari- genetic value using genome-wide dense marker maps. Genetics ants in 3000 rice genomes. Genome Res 29:870–880. https://​ 157:1819–1829. https://​doi.​org/​10.​1093/​genet​ics/​157.4.​1819 doi.​org/​10.​1101/​gr.​241240.​118 Nachimuthu VV, Raveendran M, Duraialaguraja S et al (2015) Gianola D, Foulley J (1983) Sire evaluation for ordered categorical Analysis of population structure and genetic diversity in rice data with a threshold model. Genet Sel Evol 15:201. https://​doi.​ germplasm using SSR markers: an initiative towards association org/​10.​1186/​1297-​9686-​15-2-​201 mapping of agronomic traits in Oryza sativa. Rice 8:30. https://​ Goddard ME, Hayes BJ (2007) Genomic selection. J Anim Breed doi.​org/​10.​1186/​s12284-​015-​0062-5 Genet 124:323–330. https://​doi.​org/​10.​1111/j.​1439-​0388.​2007.​ Olesen I, Perez-Enciso M, Gianola D, Thomas DL (1994) A com- 00702.x parison of normal and nonnormal mixed models for number of Haberer G, Kamal N, Bauer E et al (2020) European maize genomes lambs born in Norwegian sheep. J Anim Sci 72:1166–1173. highlight intraspecies variation in repeat and gene content. Nat https://​doi.​org/​10.​2527/​1994.​72511​66x Genet 52:950–957. https://​doi.​org/​10.​1038/​s41588-​020-​0671-9 Pérez P, de Los Campos G (2014) Genome-wide regression and pre- Habier D, Fernando RL, Kizilkaya K, Garrick DJ (2011) Extension of diction with the BGLR statistical package. Genetics 198:483– the bayesian alphabet for genomic selection. BMC Bioinformatics 495. https://​doi.​org/​10.​1534/​genet​ics.​114.​164442 12:186. https://​doi.​org/​10.​1186/​1471-​2105-​12-​186 Pérez-Enciso M, Rincón JC, Legarra A (2015) Sequence- vs. chip- Herbrich R, Graepel T, Campbell C (1999) Bayes point machines: esti- assisted genomic selection: accurate biological information mating the Bayes point in Kernel space. In: Proceedings of IJCAI is advised. Genet Sel Evol 47:43. https://​d oi.​o rg/​1 0.​1 186/​ workshop support vector machines. Stockholm, pp 23–27 s12711-​015-​0117-5 Jackson MT (1997) Conservation of rice genetic resources: the role of Purcell S, Neale B, Todd-Brown K et al (2007) PLINK: a tool set the International Rice Genebank at IRRI. Plant Mol Biol 35:61– for whole-genome association and population-based linkage 67. https://​doi.​org/​10.​1007/​978-​94-​011-​5794-0_6 analyses. Am J Hum Genet 81:559–575. https://​d oi.​o rg/​1 0.​ Jighly A, Lin Z, Pembleton LW et al (2019) Boosting genetic gain 1086/​519795 in allogamous crops via speed breeding and genomic selection. Quadrana L, Silveira AB, Caillieux E, Colot V (2021) Detection of Front Plant Sci 10:1364. https://​doi.​org/​10.​3389/​fpls.​2019.​01364 transposable element insertions in Arabidopsis using sequence Kaler AS, Purcell LC, Beissinger T, Gillman JD (2022) Genomic capture. Methods Mol Biol 2021:141–155 prediction models for traits differing in heritability for soybean, Reinoso-Peláez EL, Gianola D, González-Recio O (2022) Genome- rice, and maize. BMC Plant Biol 22:87. https://​doi.​org/​10.​1186/​ enabled prediction methods based on machine learning. s12870-​022-​03479-y Methods Mol Biol 2467:189–218. https://​d oi.​o rg/​1 0.​1 007/​ Kassambara A, Mundt F (2020) Package ‘factoextra’: extract and visu- 978-1-​0716-​2205-6_7 alize the results of multivariate data analyses. CRAN- R Package Robertsen CD, Hjortshøj RL, Janss LL (2019) Genomic selection in 84 cereal breeding. Agronomy 9:95 13 3222 Theoretical and Applied Genetics (2022) 135:3211–3222 Rosegrant MW, Cline SA (2003) Global food security: challenges Vendrell-Mir P, Barteri F, Merenciano M et al (2019) A benchmark of and policies. Science 302:1917–1919 transposon insertion detection tools using real data. Mob DNA Santiago N, Herráiz C, Ramón Goñi J et al (2002) Genome-wide 10:53. https://​doi.​org/​10.​1186/​s13100-​019-​0197-9 analysis of the emigrant family of MITEs of Arabidopsis thali- Visscher PM, Yengo L, Cox NJ, Wray NR (2021) Discovery and ana. Mol Biol Evol 19:2285–2293. https://​doi.​org/​10.​1093/​oxfor​ implications of polygenicity of common diseases. Science djour​nals.​molbev.​a0040​52 373:1468–1473 Slowikowski K (2020) ggrepel: automatically position non-overlap- Vu VQ (2011) ggbiplot: a ggplot2 based biplot. R package version ping text labels with “ggplot2” Wang B, Li J (2019) Understanding the molecular bases of agronomic Spindel JE, Begum H, Akdemir D et al (2016) Genome-wide pre- trait improvement in rice. Plant Cell 31:1416–1417 diction models that incorporate de novo GWAS are a powerful Wheelan SJ, Scheifele LZ, Martinez-Murillo F et al (2006) Transposon new tool for tropical rice improvement. Heredity 116:395–408. insertion site profiling chip (TIP-chip). Proc Natl Acad Sci U S https://​doi.​org/​10.​1038/​hdy.​2015.​113 A 103:17632–17637. https://​doi.​org/​10.​1073/​pnas.​06054​50103 Steranka JP, Tang Z, Grivainis M et al (2019) Transposon insertion Xu F, Sun X, Chen Y et al (2015) Rapid identification of major QTLs profiling by sequencing (TIPseq) for mapping LINE-1 insertions associated with rice grain weight and their utilization. PLoS ONE in the human genome. Mob DNA 10:8. https://​doi.​org/​10.​1186/​ 10:e0122206. https://​doi.​org/​10.​1371/​journ​al.​pone.​01222​06 s13100-​019-​0148-5 Xu Y, Wang X, Ding X et al (2018) Genomic selection of agronomic Sun C, Hu Z, Zheng T et al (2017) RPAN: rice pan-genome browser traits in hybrid rice using an NCII population. Rice 11:32. https://​ for ∼3000 rice genomes. Nucleic Acids Res 45:597–605. https://​ doi.​org/​10.​1186/​s12284-​018-​0223-4 doi.​org/​10.​1093/​nar/​gkw958 Xu Y, Liu X, Fu J et al (2020) Enhancing genetic gain through genomic Team RC (2021) R: a language and environment for statistical selection: from livestock to plants. Plant Commun 1:100005. computing v. 3.6. 1 (R Foundation for Statistical Computing, https://​doi.​org/​10.​1016/j.​xplc.​2019.​100005 Vienna, Austria, 2019). Sci Rep 11:12957 Xu Y, Ma K, Zhao Y et al (2021) Genomic selection: a breakthrough Tehseen MM, Kehel Z, Sansaloni CP et al (2021) Comparison of technology in rice breeding. Crop Journal 9:669–677 genomic prediction methods for yellow, stem, and leaf rust Zhao H, Mitra N, Kanetsky PA et al (2018) A practical approach resistance in wheat landraces from Afghanistan. Plants 10:558. to adjusting for population stratification in genome-wide asso- https://​doi.​org/​10.​3390/​plant​s1003​0558 ciation studies: principal components and propensity scores Tessema BB, Liu H, Sørensen AC et al (2020) Strategies using genomic (PCAPS). Stat Appl Genet Mol Biol. https://​doi.​org/​10.​1515/​ selection to increase genetic gain in breeding programs for wheat. sagmb-​2017-​0054 Front Genet 11:578123. https://​doi.​org/​10.​3389/​fgene.​2020.​ 578123 Publisher's Note Springer Nature remains neutral with regard to Tibshirani R (2011) Regression shrinkage and selection via the lasso: jurisdictional claims in published maps and institutional affiliations. a retrospective. J R Stat Soc Ser B Stat Methodol 73:273–282. https://​doi.​org/​10.​1111/j.​1467-​9868.​2011.​00771.x VanRaden PM (2008) Efficient methods to compute genomic predic- tions. J Dairy Sci 91:4414–4423. https://​doi.​org/​10.​3168/​jds.​ 2007-​0980 13