Comparisons Of Genomic Selection Methods And Studies Of Multivariate GBLUP Models

Genomic selection(GS)aims to use whole-genome markers to estimate the effects of all loci,and thus predict the phenotypes of unknown population,which provides a new method for animal and plant breeding.The existing GS methods mainly include RR-BLUP,GBLUP,BayesA,BayesB,BayesC?,BayesianLASSO,and so on.It is of great theoretical and practical significance to make clear the characteristics and applicable conditions of the above methods.Therefore,comparisons of these GS methods were carried out in this paper.In addition,the traditional GS approach focuses on a single phenotypic trait based on single-environment models,ignoring genetic correlation between traits and environmental correlation between environments.Moreover,most of them only consider the simplest additive effects,and can't effectively estimate the dominance and other non-additive effects.In this study,multivariate(MV)GS models with dominance effects were developed to perform multi-trait or multi-environment analyses to predict target traits more effectively.In addition,we carried out the studies on GS methods based on selection index,simultaneously predicting and screening multiple traits of crops to achieve more comprehensive and reliable selection.The main contents of this paper contains the following four parts:1.Comparisons of genomic selection methodsIn this study,six GS methods,including RR-BLUP,GBLUP,BayesA,BayesB,BayesC?and Bayesian LASSO were compared regarding inference under different conditions,using real data from a wheat data set and simulated scenarios with different number of QTL and different heritability.Two types of simulation were performed in this research.First,with an assumed heritability of 0.5,true breeding values were simulated based on 20,60,180 and 540 QTL.Second,after setting the QTL number to 20,simulations were varied with heritabilities of 0.3,0.5 and 0.7.This study showed that the genetic architecture of a trait should be fully considered when a GS method is chosen.The selective shrinkage models were sensitive to the number of QTL,while RR-BLUP and GBLUP showed strong robustness.If a small amount of loci(20)had a large effect on a trait,great differences were found between the predictive ability of various methods.BayesC? and BayesB were recommended in this scenario.If a trait was controlled by more genes(60,180),the absolute differences between the various methods were small,but BayesA and Bayesian LASSO were also found to be slightly superior to the other methods.If a trait was controlled by an extreme number of minor genes(540),the differences between the various methods were very small.But the results of simulation analysis and the real wheat yield prediction showed that RR-BLUP and GBLUP were more suitable for the trait controlled by a large amount of QTL.2.Predicting rice hybrid performance using GBLUP models based on North Carolina mating design ?The present study investigated the predictive ability of genomic best linear unbiased prediction(GBLUP)models for rice hybrids based on the North Carolina mating design II(NCII),in which a total of 115 inbred rice lines were crossed with five male sterile lines.Using eight traits of the 575(115×5)hybrids,including grain yield per plant(GY),thousand-grain weight(TGW),productive panicle number per plant(PN),plant height(PH),primary branch number(PB),secondary branch number(SB),grain number per panicle(GN)and panicle length(PL),univariate(UV)GBLUP models were used to predict hybrid peformance.A univariate(UV)GBLUP model including only additive effects is called UV-A,whereas a univariate GBLUP model including additive and dominance effects is called UV-AD.The prediction results of cross-validation indicated that each trait is primarily controlled through additive variance.However,pairwise comparisons illustrated that predictive ability of UV-AD was significantly higher than that of UV-A for PH,PB,SB,and GN,showing that for these traits,including dominance effects could improve the predictive ability.For GY,TGW,PN and PL,there were no significant differences between the predictive ability of UV-A and UV-AD.In each cycle of the cross-validation,we sorted the predicted phenotypic values in descending order and selected different number of top crosses to observe the benefits of prediction.The results showed that the average selection advantage of top populations was not directly related to the heritability of traits.For a low-heritability trait,such as GY,a modest increase in the number of top selection could generate a stable higher mean phenotypic value.Using UV-A,the predicted phenotypic values of GY for the 6555 crosses between the 115 inbred lines were sorted in descending order,and the mean predicted GY of the top 100 crosses is 51.78±1.38,which was much higher than the average predicted GY of the entire hybrid population(38.94).Additionally,performance of the potential crosses between the five male sterile lines and the 3023 rice varieties in the 3000-rice genomes project was predicted.Using UV-A,the mean predicted GY of the top 100 crosses was 44.43±0.52,which was much higher than the average predicted GY of the entire hybrid population(38.50).The results provided a new reference path for hybrid breeding of rice and other crops using GS methods.3.Studies on multivariate GBLUP modelsBiological traits are most likely associated.Simultaneously modeling multiple traits can make use of genetic and environmental correlated information,improving the accuracy of prediction effectively.Using the NCII rice data set mentioned previously,this study investigated multi-trait and multi-environment GBLUP models.The general multivaritate model with only additive effects was expanded,leading to the multivariate model MV-AD including additive and dominance effects,and the multivariate model MV-ADE including additive effects,dominance effects and common environmental effects.In addition,a new efficient multivariate model(MV-ADV)was developed utilizing a multivariate relationship matrix constructed with auxilary variates.The prediction results showed that the performance of MV-ADE and MV-ADV were superior to MV-AD and UV-AD.Although the absolute advantage of MV-ADV over MV-ADE was marginal,pairwise comparisons showed that the predictive ability of MV-ADV was significantly higher than that of MV-ADE.A multi-trait model is expected to increase the accuracy of prediction by incorporating information from genetically correlated traits,especially in the multi-trait scenario for a low-heritability target trait.For a high-heritability trait like TGW,the performance of multi-trait models was not improved significantly,and thus univariate prediction is a good choice.Although the joint predictive ability of ME was better than that of SE,it had less advantages compared with MT prediction,showing that the environmental difference weakened the benefits of multivariate prediction.4.Studies on GS methods using selection indexSelection index(SI)can be used to construct a comprehensive index for the comprehensive selection of traits by using genetic information between traits.In this study,using a rice data set,combined with a large number of simulation designs,selection index was established by making use of multiple auxiliary traits correlated to the target trait.Then we constructed indices such as predictive ability of SI,SI-direct accuracy,SI-direct predictive ability,SI-assisted accuracy and SI-assisted predictive ability.And thus a new method of GS based on SI was explored.The cross-validation results showed that the method can greatly utilize the genetic information of the target trait from the auxiliary traits,and construct a selection index to achieve SI-direct or SI-assisted prediction of the target trait.The higher the genetic correlation between the auxiliary traits and the target trait,the higher the predictive ability of SI,the SI-direct accuracy and the SI-direct predictive ability were.In most cases,the SI-direct accuracy couldn't surpass the accuracy of GS,but could be very close to this level.The SI-assisted prediction method could achieve higher accuracy than the general GS prediction for the target trait.The higher the genetic correlation between the auxiliary trait and the target trait,the higher the SI-assisted accuracy and the SI-assisted predictive ability were.