Abstract: With important progress in marker technologies, marker-assisted selection (MAS) has been used broadly for the crop improvement. Biparental populations are designed for the detection of quantitative trait loci (QTLs), but their application is retarded. The association mapping (AM) is applied directly to natural populations, which has been proposed to mitigate the lack of relevance of biparental populations in QTL identification. Many QTLs are identified by the two methods, which have encouraged genetic improvement of crop. However, they are using significant thresholds to identify QTL from estimated means that estimated effects are biased. Therefore, small-effect QTLs can’t be identified and missed entirely, while lots of traits of crop are controlled by those small-effect QTLs. Genomic selection (GS) has been proposed to make good for these deficiencies. Genomic selection predicts the breeding values of lines in a population by analyzing their phenotypes and high-density marker scores, and by including all markers in the model, and benefits from unbiased estimation of all chromosome segment effects, even when they are small. The GS incorporates all marker information in the prediction model, which avoids biased marker effect estimates and captures more of the variation from small-effect QTLs. Furthermore, markers carry information on the relatedness among lines, which contributes to prediction accuracy. Such accuracies are sufficient to select parents strictly on the basis of marker scores even for traits such as yield, tolerance to abiotic stress. From the view of perspective of the products from plant breeding, the genomic selection would greatly accelerate the breeding cycle, and enhance annual gains. GS would develop a prediction model from training population, genotyped and phenotyped, by estimated the markers effects. Then GS model would take genotypic data from candidate population to predict genomic estimated breeding values (GEBV), and there are some methods used for GS model, such as best linear unbiased prediction (BLUP), ridge regression BLUP (RR-BLUP),and Bayesian linear regression (BLR). These models are well used for crop genomic selection breeding. However, all the models are developed based on line or regression, while the relationships of genetic sites in life are not non-line or regression. The neural network was introduced to genomic selection in crop improvement in this study. The crop genomic selection model was optimized by non-linear model system. Therefore, the high efficient genomic selection system was established, and the prediction results were compared with these of other linear models, such as RR-BLUP, BLR.In wheat genetic data simulation, the correlation coefficient between the true breeding value of unphenotyped experimental lines and that predicted by genomic selection based on the neural network reached 0.6636, while that of RR-BLUP, BLR and BLR with pedigree information was 0.6422, 0.6294, and 0.6573, respectively. Meanwhile, the best prediction was 0.8379, which indicated the genomic selection based on the neural network is superior to these of other linear regression models. This level of accuracy was sufficient for selecting for agronomic performance using marker information alone. Such selection would substantially accelerate the breeding cycle, and enhance gains per unit time. Therefore, this research showed that GS has more potential for incorporating it into breeding schemes.

Key words: Genomic selection, Wheat, Artificial neural network, Ridge regression BLUP, Bayesian linear regression