Harnessing Deep Learning in Searching Wild Relatives of Domestic Animals

Wild relatives of domestic animals are crucial reservoirs of genetic diversity, yet pervasive hybridization with domestic animals poses significant conservation challenges. Here, we developed a deep learning-based pipeline, consisting of a multi-layer perceptron for SNP panel selection and a Deep & Cross Network for model training, to discern wild relatives from their closely related domestic animals using genomic SNP data. Leveraging the 1960 genomes from 164 red jungle fowl (RJF; Gallus gallus) and 1796 domestic chicken samples, we applied this pipeline to yield the RJF identification model based on a 285-SNP panel. We employed this model to characterize domestic chickens, RJF, and hybrids in the independent genomic datasets from contemporary samples and historical specimens, respectively. The accuracy was 97.8% for historical samples with missing genotypes. The benchmarking multiple hybrid detection tools indicated that the RJF identification model was effective and practical. The further application to the genomic data from wild boar (Sus scrofa), domestic pigs, and their hybrids validated the pipeline. Our method has potential in not only monitoring genetic diversity in wild relatives of domestic animals but also supporting animal genetic resource conservation and management.