First Report of Recurrent Genomic Selection with Real Data in Popcorn and Genetic Gain Increases

Recurrent Selection increases the frequencies of favorable alleles for economically important traits, which in the case of popcorn are popping expansion and grain yield. However, is time-consuming, since each selection cycle consists of three stages: progeny development and evaluation, and recombination of the best families. With the Recurrent Genomic Selection use, the time required for each selection cycle can be shortened, as it allows the evaluation and recombination phases to be performed simultaneously, reducing the time needed to complete one selection cycle to only one growing season. In this respect, the objective of this study was to determine the selection accuracy and genetic gains for different selection strategies: PhEN = estimates based exclusively on the phenotypic data of 98 plants; PhEN + GEN = estimates based exclusively on the phenotypic and genotypic data of 98 plants; and GEN = estimates based exclusively on SNP marker genotyping. The following traits were evaluated: 100-grain weight, ear height, grain yield, popping expansion, plant height, and popcorn volume. Field trials were carried out with 98 S1 progenies, at two locations, in an incomplete block design with three replications. The parents of these progenies were genotyped with a panel of ~ 21K SNPs. From the results based on the predictions by strategy GEN, at different selection intensities, the average annual genetic gain for the different traits was 29.1% and 25.2% higher than that by the strategies PhEN and GEN + PhEN for 98 selection candidates; 148.3% and 140.9% higher for 500; and 187.9% and 179.4% higher for 1,000 selection candidates, respectively. Therefore, recurrent genomic selection may result in a high genetic gain, provided that: i) phenotyping is accurate; ii) selection intensity is explored by genotyping several plants, increasing the number of selection candidates, and iii) genomic selection is used for early selection in recurrent selection.

22 Consequently, the time required for one selection cycle is shortened to only one growing season.

23
In the literature, no studies on GS in real popcorn breeding populations have assessed the 24 genetic gains for agronomic traits and the viability of this technique in comparison with traditional 25 breeding techniques. However, GS has been proved to be efficient to increase selection gains per time 26 unit and as a tool to predict effects of general combining ability of testcrosses and genotypic values 27 of double-haploid maize lines (5); in the breeding of allogamous lines (6) and in the prediction of 28 genotypic values of single-cross maize, wheat and rice hybrids (7,8).

29
In some cases, GS also proved to be more efficient per unit of time than phenotypic or marker-30 assisted selection (9). In addition, other studies also reported that GS is efficient and can be used in 31 plant breeding (10)(11)(12).
32 Therefore, the objective of this study was to determine the selection accuracy for different 33 selection strategies and the respective genetic gains at different selection intensities, with a view to 34 assessing the efficiency of recurrent genomic selection in popcorn breeding.
22 stability (17,18).   27 based only on marker data. This model fitting process was repeated 10 times, and in each cycle, the 28 genetic merit of a different group of individuals was predicted with the ignored phenotype. This 29 process is known as ten-fold cross-validation (23,24), and was used in this study to test the efficacy 30 of early selection, using a previously fitted model. 6 1 By the strategy GEN, predictions for different selection intensities were tested, for 98 to 1,000 2 selection candidates, setting an absolute value of 40 selected plants. Thereafter, a comparative 3 evaluation of the annual genetic gain by the strategies GEN, PhEN and GEN + PhEN was carried out.
4 Accuracy was estimated as proposed by (25) considering only the estimates of the 98 individuals that 5 passed the filter for the loss rate of SNP data in this study. 6 The following formula was used to compute the genetic gain:

15
In this study, the mean-based heritability was medium to high for most traits, and was < 0.5 16 only for PV and > 0.8 for 100GW, EH and PH (Table 1)

31
In conventional recurrent selection, selection is based on exclusively phenotypic data; in this 32 study, when including information from the set of 10,507 SNPs markers, accuracy increased slightly 8 1 for most traits. In this way, the model GEN + PhEN was in the mean ≈2.5% more accurate than the 2 traditional model, which ignores the markers (PhEN). The accuracy based exclusively on the strategy 3 GEN was between 0.25 and 0.39, i.e., much lower than the values based on PhEN and GEN + PhEN 4 (Table 1).

5
The accuracy values for the different strategies were highest for the traits related to crop

30
Genetic gain is the amount of increase in performance that is achieved in breeding programs 31 over consecutive selection cycles (30). The selection gain depends on the selection differential, which 32 in turn is the difference between the mean of the selected group and the mean of the original 33 population. Therefore, in the selection process, the higher the selection pressure, the greater the 34 differential, and consequently, the higher the genetic gain (2).  (Figure 3). The strategy GEN obtained significant annual gains for the most important agronomic traits 14 GY and PE, which were 32.8% and 28.2% higher for 98 candidates, and 155.4% and 147.1% higher 15 for 500 candidates, respectively, than by strategy PhEN. By strategy GEN, the average annual gain 16 increased 92.4% up to 500 selection candidates, and thereafter, the selection gain kept on increasing, 17 although to a lesser extent, since for 500 to 1,000 selection candidates, the gain was 15.9% ( Figure   18 3).

27
The variance of the significant genotype-location interaction found for the traits GY and PV 28 is a result of differentiated responses of the tested genotypes in different environments. However, to 29 infer whether the interaction is significant is only a qualitative approach, and the correlated response 30 can only be investigated by the genetic correlation between the environments (34). For the trait GY, 31 with a high genetic correlation, the selection in one environment will imply in a positive response in 32 another. As the genetic correlation between the environments was positive, the selection based on the 10 1 general additive value, i.e., for both environments, will provide a positive selection response for both 2 environments. 3 The heritability of PE (0.62) exceeded that of GY (0.51). Also (35) observed a lower 4 environmental influence on the expression of PE than on GY. The higher heritability of PE can also 5 be explained by the fact that the trait is predominantly controlled by additive genes (36-39).

6
The low heritability (0.43) obtained with PV can be explained by the fact that this trait is a 7 combination of GY and PE, also incorporating the environmental influence in the trait expression of 8 two polygenic traits. According to (40), genomic selection will provide relevant genetic gains for 9 plant breeding, even for low heritability traits, by reducing the duration of the selection cycle.

13
The superiority of the strategy GEN over PhEN and GEN + PhEN agrees with the findings of 14 (6,50,51), who suggested that the potential of GS is superior to that of phenotypic selection. However, 15 this depends on a fitted prediction model, to be used in one or more RS cycles without recalibration.
16 According to (52), the success of this strategy depends to a large extent on the maintenance of 17 significant accuracy values, when the selection candidates are separated by one or more cycles of the 18 training generation of the model.

19
According to (51), it is not possible to determine after how many cycles the model will become 20 effective, because changes in the LD pattern, allele frequencies and polymorphism losses are 21 unpredictable. According to (4), it is claimed that GS can be used from one to three RS cycles without 22 a phenotypic evaluation, although this statement will need further evaluation. On the other hand, (41) 23 performed five GS cycles and observed that the genetic gains in maize grain yield generally agreed 24 with the predicted level, although the gains after the first cycle were unstable.

25
The results of strategy PhEN + GEN were slightly better than those of PhEN, due to the fact 26 that strategy PhEN ignores the marker data to establish the kinship matrix, since for UENF-14 the 27 kinship matrix is the proper identity (G = I). The population UENF-14 is unstructured, due to the 28 recombination it underwent, with the progenies selected in the RS cycle, immediately before this 29 study. According to (44) the closer the actual genetic relationship between the training and validation 30 populations, the fewer markers will be required to achieve a satisfactory prediction accuracy.

31
The predicted annual gains for GY and PE by strategy GEN were much higher than those 32 reported in previous RS studies for this population, where selection was performed exclusively with 33 the evaluation of phenotypic family means (37,53-59).