Designed phenotyping for genomic selection

Genomic selection (GS) is increasingly used in plant breeding to predict genetic merit using genome-wide molecular markers. While advances in high-throughput genotyping and statistical modeling, including machine learning, have made GS operational across many crop species, realized prediction accuracy continues to vary substantially among traits, environments, and breeding programs. Much of the GS literature has focused on markers and models, whereas comparatively less attention has been devoted to the design of the phenotyping systems that generate the information required for model training. In this perspective, we argue that GS is built on an information-generation system in which breeding and phenotyping generate information, whereas genotyping and modeling capture, utilize, and propagate that information. We propose the concept of designed phenotyping, in which genomic prediction accuracy is constrained by two complementary factors: phenotypic information quality and information transferability. Multi-environment trial design primarily determines phenotypic information quality through its effects on entry-mean heritability, whereas training population design influences both information quality and transferability through its effects on genetic variability, connectivity, relatedness, and relevance. Together, these factors define the practical upper limit of genomic prediction accuracy. We further argue that phenotyping for GS should not be viewed as a one-time activity but as a continuous investment in the information-generation system underlying future genetic gain. Consequently, designed phenotyping represents a strategic framework for integrating trial design and training population design to sustain long-term breeding progress.