Abstract
<title>Abstract</title> <p>High nucleotide diversity and rapid linkage disequilibrium decay erode the stability of genomic predictions across generations in hyper-diverse genomes. Whether any variant classes can maintain associations that are sufficiently stable to support prediction across generations remains unresolved. Aquaculture breeding programs provide a unique experimental setting to quantify transmitted genetic effects through common-garden rearing and large-scale progeny validation. Leveraging this design in the Pacific oyster as a hyper-diverse model, we couple genome-wide variant discovery with cross-generational validation to quantify transmitted additive effects under minimized environmental variation. We uncover a fundamental decoupling between heritability and transmitted prediction accuracy. Rare variants in weak-LD regions explain substantial heritable variation, yet their effects are evolutionarily unstable and translate poorly into prediction accuracy across generations. In contrast, ancient variants with high allele frequencies and strong LD preserve persistent haplotype structure and dominate prediction accuracy. Prioritizing these evolutionarily persistent variants substantially improves accuracy relative to whole-genome and functionally annotated models and yields consistent gains when independently validated in representative fish and shrimp species. These findings establish an evolution-informed framework for genomic prediction in hyper-diverse genomes, in which prediction accuracy is governed by the evolutionary persistence of variants rather than by the amount of genetic variance they explain.</p>