Study And Application Of Genomic Selection For Predicting Hybrid Performance In Rice

The utilization of rice heterosis is an important approach to increase rice yield,and parental selection is the key and difficult issue in the hybrid breeding.Currently,the parental selection mainly relies on the experience of breeders,which has great uncertainty and takes a lot of effort to finish phenotypic identification in the field.Compared with the traditional breeding techniques,molecular breeding by design can realize the directional and efficient’precise breeding’ and greatly improve the breeding efficiency.The genomic selection(GS)provides opportunities for us to predict the performance of the hybrids.Theoretically,he phenotypic values of all hybrids can be predicted using their parental genome-wide markers and a small fraction of hybrids for field evaluation.The success of GS depends on the accuracy of phenotypic prediction.In order to improve the accuracy of hybrid prediction.In order to improve the accuracy of rice hybrid phenotype prediction,this study utilizes the SNPs of whole genome of 120 rice varieties by sequencing and the phenotype of eight traits of 575(115×5)hybrids to carry out a comparative study of six GS methods,including genomic best linear unbiased prediction(GBLUP),least absolute shrinkage and selection operation(LASSO),Bayes B,partial least square(PLS),support vector machine(SVM)and reproducing kernel Hilbert space(RKHS).We investigate the influence of various methods on different traits and the effect of different marker number on the predictability of trait.At last,we use the data of whole genome SNP of 3023 rice resources from ’3K rice genomic project’,to predict the phenotype of all hybrids derived from 120 rice varieties each mating with 3023 rice accessions,in which 575 hybrids are selected as a traing sample.The results are as follows:1)We first obtained the 2561889 SNP markers from the 120 rice varieties and the 6572189 SNP markers from 3023 rice accessions by referring to the Nipponbare genome.Then we got 2395866 SNP markers(all SNP)which locate at the same position in the rice genome by comparing the two genome datasets.At the same time,we compared the 2561889 SNP markers with the 996009 SNP markers which had been filtered according to LD in the 3K rice genomic project,and obtained 116482 SNP markers(core SNP)shared by the two datasets.We found the GBLUP predictive accuracy was nearly the same by using 116482 SNP markers and 2395866 SNP markers.To speed up the operation we compared various GS methods by using the 116482 SNP markers.Meanwhile we predicted by GBLUP with the all SNP firstly,and then we reduced the number of the all SNP continuously,at last we found that the change of marker number had no effect on the predictive accuracy if the number of SNP markers was more than 2K.2)The results showed that the predictability of different methods had very significant difference.In the six methods,GBLUP and LASSO had a better predictability,SVM and PLS had the worst predictability.In addition,the predictability of different traits had very significant differences,the predictability of thousand kernel weight was the highest,grain yield and panicle number per plant had the lowest predictability in the eight traits.It is obvious that the higher heritability of the trait and the higher predictability.Furthermore,we also found different methods might adapt to different traits.3)According to the prediction,in the 120×3023 possible hybrids,the average grain yield of top 100 hybrids increased 35.5%compared to the average grain field of the whole hybrids,and thousand kernel weight,panicle number per plant,plant height,primary branch number,secondary branch number,grain number per panicle and panicle length increased by 22.4%,29.7%,43.4%,33.1%,59.7%,85.3%and 19%,respectively.The results of this study can be helpful to breed rice hybrids by GS.