How accurate is genomic prediction across wild populations?

Abstract

Evolutionary ecology seeks to understand causes and consequences of evolutionary changes across time and space, and genomic data present novel opportunities to investigate these processes. Genomic prediction—predicting individual genetic values from high-density marker data—has revolutionized breeding programs and medical genetics. In wild populations, however, genomic prediction has been used in comparatively few studies, and largely within populations. Applications that instead operate across populations could answer questions related to spatially varying evolutionary processes, such as local adaptation. A severe challenge for across-population genomic prediction, however, is the decrease in accuracy when training models on data from one population and predicting genetic values in another. Here, we applied genomic prediction across wild house sparrow populations and compared the accuracy to within-population models. We also highlighted limitations of the current theory for genomic prediction accuracy, and sought to provide a mechanistic understanding of the across-population accuracy by relating it to several population-differentiation measures. Predictions across populations were generally less accurate and more variable than within populations, and across-population accuracy covaried with some population-differentiation metrics. Our results underline the necessity of understanding the mechanisms governing genomic prediction accuracy, and of developing methods that exploit genomic data in novel ways.