Phenomic selection is competitive with genomic selection for breeding of complex traits

Abstract

Abstract The efficiency of breeding programs depends on the ability to screen large numbers of individuals. For complex traits like yield, this can be assisted by genomic selection, which is based on estimating breeding values with genome‐wide marker data. Here, we evaluate phenomic prediction, which, similar to its genomic counterpart, aims to predict the performance of untested individuals but using near‐infrared spectroscopy (NIRS) data. In a large panel of 944 soybean [Glycine max (L.) Merr.] recombinant inbred lines phenotyped for seed yield, thousand‐seed weight, and plant height at three locations, we demonstrate that the phenomic predictive abilities are high and comparable with those obtained by genomic prediction. We found that ridge regression best linear unbiased prediction performs well for phenomic prediction and that the number of wavelengths can be reduced without a decrease in predictive ability. For prediction at different locations, NIRS data from a single location can be used. However, NIRS data from different environments, like years, should be connected by common genotypes in training and prediction sets. Phenomic prediction appears to be less susceptible to relatedness between individuals in training and prediction sets than genomic prediction, as generally half‐sib but also unrelated families achieved high predictive abilities. Moreover, for the same training set sizes phenomic prediction resulted in higher predictive abilities compared to genomic prediction. Phenomic prediction can be applied at different stages in a breeding program, and collectively our results highlight the potential of this approach to increase genetic gain in plant breeding.