Predicting Geographic Location from Genetic Variation with Deep Neural Networks

Abstract

6Most organisms are more closely related to nearby than distant members of their species, creating 7 spatial autocorrelations in genetic data. This allows us to predict the location of origin of a genetic 8 sample by comparing it to a set of samples of known geographic origin. Here we describe a deep 9 learning method, which we call Locator, to accomplish this task faster and more accurately than 10 existing approaches. In simulations, Locator infers sample location to within 4.1 generations of 11 dispersal and runs at least an order of magnitude faster than a recent model-based approach. 12 We leverage Locator's computational efficiency to predict locations separately in windows across 13 the genome, which allows us to both quantify uncertainty and describe the mosaic ancestry and 14 patterns of geographic mixing that characterize many populations. Applied to whole-genome 15 sequence data from Plasmodium parasites, Anopheles mosquitoes, and global human populations, 16 this approach yields median test errors of 16.9km, 5.7km, and 85km, respectively. 17 Introduction 18In natural populations, local mate selection and dispersal create correlations between geographic 19 location and genetic variation -each individual's genome is a mosaic of material inherited from 20 recent ancestors that are usually geographically nearby. Given a set of genotyped individuals of 21 known geographic provenance, it is therefore possible to predict the location of new samples from 22 genetic information alone (Guillot et al., 2015; Yang et al., 2012; Wasser et al., 2004; Rañola et al., 23 2014; Bhaskar et al., 2016; Baran et al., 2013). This task has forensic applications -for example, 24 estimating the location of trafficked elephant ivory as in Wasser et al. (2004) -and also offers a way 25 to analyze variation in geographic ancestry without assuming the existence of discrete ancestral 26 populations. 27The most common approaches to estimating sample locations are based on unsupervised geno-28 type clustering or dimensionality reduction techniques. Genetic data from samples of both known 29 and unknown origin are jointly analyzed, and unknown samples are assigned to the location of 30 known individuals with which they share a genotype cluster or region of PC space (Breidenbach 31 et al., 2019; Battey et al., 2018; Cong et al., 2019). However, these methods require an additional 32 mapping from genotype clusters or PC space to geography, and can produce nonsensical results if 33 unknown samples are hybrids or do not originate from any of the sampled reference populations. 34Existing methods for estimating sample location that explicitly model continuous landscapes 35 use a two-step procedure. A smoothed map describing variation in allele frequencies over space 36 1 Battey et al. 2019Locator is first estimated for each allele based on the genotypes of individuals with known locations, and 37 locations of new samples are then predicted by maximizing the likelihood of observing a given 38 combination of alleles at the p...

The origins and spread of domestic horses from the Western Eurasian steppes

“…The tarpan horse came about following admixture between horses native to Europe (modelled as having 28.8-34.2% and 32.2-33.2% CWC ancestry in Ori-entAGraph 19 and qpAdm 17 , respectively) and horses closely related to DOM2. This is consistent with LOCATOR 20 predicting ancestors in western Ukraine (Fig 3c) and refutes previous hypotheses depicting tarpans as the wild ancestor or a feral version of DOM2, or a hybrid with Przewalski's horses 34 .…”

“…This eliminates the possibility of DOM2 ancestors further west than C-PONT and the Dnieper steppes. Furthermore, patterns of spatial autocorrelations in the genetic data 20 indicated Western Eurasia steppes as the most likely geographic location of DOM2 ancestors (Fig. 3c).…”

Chromosome‐scale inference of hybrid speciation and admixture with convolutional neural networks

“…Exploring more ways to extract information from genomic data using deep learning is likely to continue being an active area of research not only in terms of designing neural network architectures but also for thinking about how to represent genetic variation. The way that we represent genomic data as an image differs from other approaches that use the actual nucleotide alignment or genotype matrix as their input for CNN training (Battey et al, 2020;Flagel et al 2018;Suvorov et al 2019). We chose our data representation because we believed it to be an intuitive summary of pairwise coalescence times between species organized by the pattern of divergence in the underlying phylogeny.…”

Spatial population genetics in heavily managed species: Separating patterns of historical translocation from contemporary gene flow in white-tailed deer

“…This could be used, for example, to ascertain the geographic origin of poached individuals, or to estimate post-natal dispersal. To this end, we used the novel, deep-learning method LOCATOR (Battey, Ralph, & Kern, 2020) to predict the geographic origin of samples without relying upon explicit assumptions about population genetic processes underlying spatial genetic differentiation (Bradburd & Ralph, 2019). This analysis was performed iteratively across each individual, using the remaining samples to train the LOCATOR classifier, with 100 bootstrap pseudo-replicates to assess variance in geolocation.…”