Motivation Modern biobanks provide numerous high-resolution genomic sequences of diverse populations. These datasets enable a better understanding of genotype-phenotype interactions with genome-wide association studies (GWAS) and power a new personalized precision medicine with polygenic risk scores (PRS). In order to account for diverse and admixed populations, new algorithmic tools are needed in order to properly capture the genetic composition of populations. Here we explore deep learning techniques, namely variational autoencoders (VAEs), to process genomic data from a population perspective. We hope this work will encourage the adoption of deep neural networks in the population genetics community.

Results In this paper, we show the power of VAEs for a variety of tasks relating to the interpretation, classification, simulation, and compression of genomic data with several worldwide whole genome datasets from both humans and canids and evaluate the performance of the proposed applications with and without ancestry conditioning. The unsupervised setting of autoencoders allows for the detection and learning of granular population structure and inferring of informative latent factors. The learned latent spaces of VAEs are able to capture and represent differentiated Gaussian-like clusters of samples with similar genetic composition on a fine-scale from single nucleotide polymorphisms (SNPs), enabling applications in dimensionality reduction, data simulation, and imputation. These individual genotype sequences can then be decomposed into latent representations and reconstruction errors (residuals) which provide a sparse representation useful for lossless compression. We show that different population groups have differentiated compression ratios and classification accuracies. Additionally, we analyze the entropy of the SNP data, its effect on compression across populations, its relation to historical migrations, and we show how to introduce autoencoders into existing compression pipelines.