Why does genomewide prediction become ineffective after several cycles of recurrent selection?

Genomewide selection is effective if its prediction accuracy (r_MG) is high. The r_MG is known to decrease after several cycles of selection, but a systematic analysis of the factors that contribute to the decline in r_MG has not been reported. My objective was to assess what factors contribute the most to the decay in r_MG during genomewide selection. Ten cycles of genomewide recurrent selection with different genetic models were simulated for a maize (Zea mays L.) biparental cross. In the benchmark model, which involved 250 quantitative trait loci (QTLs), N = 200 plants in each cycle, and the best N_Sel = 10 plants selected in each cycle, the r_MG declined from 0.77 in Cycle 0 to 0.16 in Cycle 10. Results for truncation versus random selection indicated that directional selection itself accounted for >50% of the variation in r_MG. The decay in linkage disequilibrium across cycles of selection accounted for nearly 30% of the variation in r_MG. Genetic drift, number of QTLs, and having functional versus random markers had nonsignificant effects on r_MG. Suppression of crossing-over along with random selection maintained r_MG at 0.76–0.77 across all 10 cycles but, as expected, led to no selection gain. Because selection and a decay in linkage disequilibrium are inherent in genomewide recurrent selection, a decrease in r_MG is an inevitable price to pay for genetic gain. A new prediction model is then needed after several cycles of selection.