Genomic prediction with epistasis models: on the marker-coding-dependent performance of the extended GBLUP and properties of the categorical epistasis model (CE)

Table 2 Predictive abilities of the models on the simulated data. Comparison of the predictive abilities in terms of correlations between the measured phenotypes and the predictions for the individuals of the test sets (“Pearson’s correlation”; 100 test set genotypes were drawn randomly from all 1000 genotypes; 200 repeats for each simulated population; 20 independent simulations of population and phenotypes). Traits of different genetic architecture (additive A, dominant D, Epistasis E) and increasing number of QTL. Model abbreviations as introduced in the text. For EGBLUP, only the matrix based on the interactions was considered here

	GBLUP	EGBLUP 0,1,2	EGBLUP -2,-1,0	EGBLUP -1,0,1	EGBLUP VR	CM	CE	K
A1	0.551 ± 0.005	0.552 ±0.005	0.552 ±0.005	0.550 ± 0.005	0.372 ± 0.006	0.489 ± 0.005	0.494 ± 0.005	0.530 ± 0.005
A2	0.549 ± 0.005	0.550 ±0.005	0.550 ±0.005	0.548 ± 0.005	0.351 ± 0.006	0.486 ± 0.005	0.490 ± 0.005	0.527 ± 0.005
A3	0.569 ± 0.005	0.570 ±0.005	0.570 ±0.005	0.568 ± 0.005	0.372 ± 0.006	0.500 ± 0.005	0.504 ± 0.005	0.545 ± 0.005
D1	0.159 ± 0.006	0.160 ± 0.006	0.159 ± 0.006	0.161 ± 0.007	0.111 ± 0.007	0.174 ± 0.006	0.175 ±0.006	0.162 ± 0.006
D2	0.172 ± 0.006	0.172 ± 0.006	0.172 ± 0.006	0.171 ± 0.006	0.103 ± 0.006	0.186 ±0.006	0.186 ±0.006	0.170 ± 0.006
D3	0.156 ± 0.006	0.156 ± 0.006	0.156 ± 0.006	0.158 ± 0.006	0.116 ± 0.006	0.177 ± 0.006	0.179 ±0.006	0.160 ± 0.006
E1	0.244 ± 0.006	0.244 ± 0.006	0.244 ± 0.006	0.244 ± 0.006	0.159 ± 0.006	0.258 ±0.006	0.258 ±0.006	0.243 ± 0.006
E2	0.275 ± 0.006	0.276 ± 0.006	0.276 ± 0.006	0.277 ± 0.006	0.188 ± 0.006	0.301 ± 0.006	0.302 ±0.006	0.277 ± 0.006
E3	0.279 ± 0.006	0.278 ± 0.006	0.279 ± 0.006	0.278 ± 0.006	0.176 ± 0.006	0.304 ±0.006	0.304 ±0.006	0.276 ± 0.006

EGBLUP VR denotes the interaction model based on the by allele frequencies standardized matrix. The given values represent the empirical mean and the corresponding mean standard error across the 20 independently simulated data sets. The highest predictive ability is bold

ISSN: 1471-2105