Deep learning reveals the complex genetic architecture of a highly polymorphic sexual trait

The extraordinary variation in male guppy coloration has proven a powerful model for studying the interplay of natural and sexual selection. However, this variation has hampered the high-resolution characterization and determination of the genetic architecture underlying male guppy color, as well as clouded our understanding of how this exceptional level of diversity is maintained. Here we identify the heritability and genetic basis of male color variation using convolutional neural networks for high-resolution phenotyping coupled with selection experiments, controlled pedigrees and whole-genome resequencing for a Genome Wide Association Study (GWAS) of color. Our phenotypic and genomic results converge to show that color patterning in guppies is a combination of many heritable features, each with a largely independent genetic architecture spanning the entire genome. Autosomally-inherited ornaments are polygenic, with significant contributions from loci involved in neural crest cell migration. Unusually, our GWAS results suggest that gene duplicates from the autosomes to the Y chromosome are responsible for much of the sex-linked variation in color in guppies, providing a potential mechanism for the maintenance of variation of this classic model trait.