We first evaluated Locator’s performance in simulations of populations evolving in continuous space with varying rates of dispersal – an idealized setting in which all alleles should vary smoothly over the map. In Figure 2 we show that validation error increases along with the dispersal rate of the population. Interestingly, error is roughly constant when measured in terms of the mean per-generation dispersal rate, ranging from 3.16 to 4.09 generations of dispersal given our largest training dataset (450 samples, 100,000 SNPs; Figure 2—figure supplement 1; Supplementary file 1). This suggests that error primarily reflects the underlying biological processes of dispersal and mate selection rather than simple noise from model fitting.

Increasing the number of training samples or the number of SNPs improves accuracy for all simulations (Figure 2B). However, we observed diminishing returns when using over 10,000 SNPs or over 200 training samples. Median error for all simulations was also below 10 generations of dispersal for all but the least-dispersive simulation using just 25 training samples; suggesting that even relatively small training datasets can allow inference of broad-scale spatial locations. We discuss theoretical limits on the accuracy of genetic location estimation in Appendix 1.

We were interested to compare the performance of Locator to that of SPASIBA (Guillot et al., 2016), the current state-of-the-art method for geographic prediction of genotype data (Figure 3; Supplementary file 2). However, we were unable to succesfully run SPASIBA with 100,000 or more SNPs from a simulated dataset or on simulations with dispersal rates of 0.63 or 1.29 map units/generation, due to out-of-memory errors on a 64-bit system with 400 Gb of RAM. We could, however, compare at smaller numbers of SNPs and reduced dispersal. At a mean dispersal distance of 0.45 map units SPASIBA’s median test error was slightly lower when run on 1000 SNPs (Wilcoxon test, p=0.009) but results were similar at 100 or 10,000 SNPs. (Wilcoxon test, p=0.184 and 0.936). However, Locator is much faster – training on 10,000 SNPs in less than two minutes while SPASIBA requires around six and a half hours (Figure 2). These long run times are caused in part by the large number of training localities in our simulated data, because SPASIBA’s run time scales with the product of the number of genetic variants and the number of training localities (Guillot et al., 2016).

While the simulations conform well to modeling assumptions of most methods, we can also compare performance on empirical data. By way of example, we applied Locator and SPASIBA to subsets of SNPs from the first five million base pairs of chromosome 2L from the Ag1000G dataset (Miles and Harding, 2017; Figure 3). Locator achieves much lower mean error on all runs with more than 100 SNPs, and runs from 3.1x to 532x faster, depending on the number of SNPs. Two factors likely explain this improved accuracy: Locator can handle abrupt changes in allele frequencies across the landscape better than SPASIBA’s geostatistical model, and the limited number of sampling localities in the AG1000G dataset may act as a prior encouraging Locator to assign samples to specific sampling sites. Maps of predictions from both methods are shown in Figure 3—figure supplement 1. Extrapolating from these run times, running a windowed whole-genome analysis of Anopheles in SPASIBA would require roughly 70 days of computation on an 80-CPU system for model training alone, versus 3.2 hr on one GPU for Locator.

By running Locator in windows across the genome we aim to integrate over error associated with the model training procedure while also representing the inherent uncertainty caused by spatial drift of ancestral lineages backwards in time (Kelleher et al., 2016). This produces a cloud of predicted locations distributed around the true sample location (Figure 4). For individuals near the center of the landscape, these clouds are roughly symmetrical, as expected from our model. Predictions for individuals close to the edge of the landscape appear slightly asymmetrical and are bounded by the true landscape edges, suggesting that our networks have learned the rough shape of the sampled range. The true location was within the 50% contour of a 2d-kernel density surface estimated from the set of per-window predictions for all test samples, demonstrating that this distribution is indeed centered on the true location. We also tested the alternate approach of bootstrapping over a single set of SNPs, which could be useful for smaller datasets or those lacking a reference alignment. Results for this method are discussed in Figure 4—figure supplement 1.

Windowed analyses for the three empirical systems we studied are shown in the bottom panels of Figures 5, 6, 7. We discuss the implications of these predictions for each species in the following sections, but in general we find that the windowed analysis provides a good depiction of uncertainty in a sample’s location – either surrounding a single location for samples with low error, or distributed across a wide region including multiple training localities for samples with high error. In several cases, predicted locations also project in the direction of known historic migrations (as in human data), or are split among localities shown in previous analyses to experience high gene flow (as in Anopheles).

We summarize genome-wide window predictions in two ways: 1) by taking a kernel density estimate of the predictions and then finding the point of maximum density, and 2) by computing the centroid of the windowed predictions. These estimates are similar in spirit to ensemble prediction methods (Ho and forests, 1995; Breiman, 1996), but should not be interpreted as true statistical confidence intervals. At least in the context of windowed analyses, differences in predictions among windows appears to primarily reflect variation in ancestry rather than uncertainty in the inference itself, so we suggest the intervals returned by Locator’s kernel density estimation are best interpreted as representing areas from which a given proportion of the genome is likely to have originated. In general, we found that the maximum kernel density estimator has lower error, but tends to show classification behavior more than the centroid estimator – snapping to a single training locality rather than interpolating between sets of localities for samples with variable window predictions.

Locator’s prediction accuracy varied widely along the human genome. To assess the sources of this variation, we first examined how recombination rate interacts with the accuracy of Locator predictions generated from different regions of the genome. A priori we might expect recombination rate to affect accuracy because in regions of the genome with higher recombination, there are a greater number of distinct genealogies, and hence a given sample has inherited from a larger subset of the possible ancestors. Test error was estimated as the distance in kilometers from the true sampling location to the geographic centroid of the cloud of per-window predictions, and is shown in Figure 8 plotted against local recombination rates from the HapMap genetic map (International HapMap Consortium, 2003). We find a relatively strong negative correlation (p<0.0001, R2=0.27) – windows with the lowest recombination rates in general have the highest prediction error. This is consistent with our expectation that regions of the genome representing a greater number of marginal genealogies will yield more accurate predictions of a sample’s location.

This finding suggests recombination-based distances may be a better basis for establishing window bounds. We tested this by replicating our analysis of HGDP samples using a 10-centiMorgan window size, which yields approximately the same number of windows across the genome. We found the distribution of centroid prediction errors across windowing methods was very similar (Figure 8—figure supplement 3) but the variance in per-window prediction error was lower when using recombination-based windows (Figure 8—figure supplement 4; Levene’s test p=0.0002, df = 654). Examining plots of window error along the genome suggests much of this effect is caused by high error in windows containing a centromere when using physical distance-based windows (Figure 8—figure supplement 1, Figure 8—figure supplement 2). Centromere presence explains 26.9% of variance in window prediction error using 10 Mb windows, but just 2.6% of variance with centiMorgan-based windows. Thus, we recommend researchers interested in analyzing individual window predictions use recombination-based windows when possible, but find that using fixed windows is sufficient for estimating individual-level locations.

Last, to examine how predictions vary near a well-known geographically differentiated region of the genome, we trained a Locator model using SNPs from within 100 kb of the EDAR gene. The rs3827760 SNP is a A→G mutation that created a derived allele which has reached high frequency in East Asian and North American populations but is rare elsewhere (Bryk et al., 2008). This variant is thought to be associated with traits including hair thickness (Fujimoto et al., 2008), and the EDAR region has been proposed as a site of recent positive selection in several analyses (Voight and Kudaravalli, 2006; Tang et al., 2007; Williamson et al., 2007). We focused on predictions for samples from central Asia, in the middle of the east-west cline in rs3827760 allele frequencies across Eurasia.

In this analysis, the direction of prediction error generally followed the genotype at rs3827760 – individuals with a G allele tend to predict east of their true location (Figure 9). This was particularly clear in heterozygous populations. For example, homozygous G/G Xibo and Hazara individuals predict east of their true location, while one homozygous A/A Uyghur sample predicts well west of the sampling site. 3 of 4 heterozygous samples also predict east of their true location, and five other A/A Central/South Asian samples predict significantly west. The discrepancies between sampling and predicted locations likely in fact represent signal: haplotypes carrying a G allele at this locus likely have more ‘close’ relatives in eastern than western Eurasia. These patterns are magnified versions of the trends seen in windowed analyses, suggesting that strong differentiation in this genomic region biases location predictions away from the center of the geographic cline.

To understand how Locator predictions vary when a sample’s true locality is not included in the training set, we ran analyses on a single window of the Anopheles data at two scales – first dropping only sites from a specific sampling location, and then dropping all sites from a given country (Figure 5—figure supplements 2–3). Prediction error is much higher for individuals from regions excluded from training – increasing from a median of 14 km when training and test samples are randomly split to 116 km when excluding individual localities, and 778 km when excluding whole countries.

In most cases, predicted locations appear to project toward the nearest locality included in the training set (Figure 5—figure supplement 3). This is particularly the case when populations at the edge of the map are excluded. Locator networks appear to learn something about the boundaries of the landscape based on the distribution of training points, and show a tendency to project towards the middle of the landscape when given a small number of SNPs (e.g. the top right panel of Figure 2A), a trivial optimization of the loss function. We also see evidence of Locator learning some nonlinear aspects of population structure in the sample. For example, when Angolan A. coluzzii are excluded from the training set many of their predicted locations project toward the A. coluzzii sample localities in Burkina Faso rather than the much closer sampling localities for A. gambiae in Cameroon and Gabon. In general, we find that Locator can interpolate unsampled localities relatively well when genetic differentiation is smooth over the landscape (as among A. gambiae localities in west Africa), but does not extrapolate outside the bounds of the training set. Sampling the full landscape, or at least a sufficient portion thereof, is thus an important consideration in running our method.

Locator transforms input data in VCF or Zarr format to vectors of allele counts per individual using the scikit-allel (Miles and Harding, 2017) and numpy (van der Walt et al., 2011) libraries. Sites with missing data are replaced with two draws from a binomial distribution with probability equal to the frequency of the derived allele across all individuals – a discrete version of the common practice of assigning missing data as the mean allele frequency in genotype PCAs (e.g. the default settings for PCA in the R package adegenet [Jombart, 2008]). We provide functions for filtering SNPs based on minor allele count, and by default remove singleton sites from the alignment prior to model fitting. The geographical x and y coordinates are scaled to have mean 0 and variance one prior to training, while allele counts are scaled prior to model fitting by a batch normalization layer within the network. Batch normalization Z-normalizes activations of a neural network during training to reduce shifts in the distribution of parameter values across batches, which allows faster learning rates and sometimes reduces overfitting (Ioffe and Szegedy, 2015).

Locator selects user-defined fraction of the samples with known locations to use in training the model (the default is 0.9); remaining samples with known locations are kept aside as ‘validation’ samples. The validation set is used to tune the learning rate of the optimizer and set the stopping time of model training, but does not directly contribute to the loss used to fit model parameters. Throughout this manuscript, we use ‘validation loss’ to refer to error estimated on the validation set, and ‘test error’ to refer to error calculated on a set of samples entirely held out from the model training procedure.

For datasets with small sample sizes the random train/test split may lead to some regions being under- or overrepresented in the training sample. To mitigate this we suggest fitting multiple models with different random seeds, yielding an ensemble of models trained on different subsets of the original dataset. Predictions from this ensemble can then be summarized in the same way as windows or bootstrap samples (see below). An example of this approach is included in the Locator documentation (https://github.com/kern-lab/locator).

We use the unphased, diploid genotype vector of each individual as input to the network, whose target output is the two-dimensional coordinates of that individual in space. Locator uses a deep neural network consisting of a stack of fully connected ‘dense’ layers, implemented using the Keras (Chollet, 2015) frontend to tensorflow (Abadi et al., 2015). Roughly speaking, the network is trained to estimate a nonlinear function mapping genotypes to locations using gradient-based optimization. Models start with randomized initial parameters and are fit to data by looping through the training set and iteratively adjusting the weights and biases of the network. We use an early stopping function to monitor loss during training and under default settings stop training runs when validation loss has not improved for 100 epochs. We also use a learning rate scheduler to decrease the learning rate of the optimizer when validation loss stops improving, which we found to be effective in preventing the trajectories of training and validation loss from diverging. The program also outputs a plot of training and validation loss after each training run (Figure 2—figure supplement 2).

Locator’s architecture uses a batch normalization layer followed by a sequence of fully connected layers with a dropout layer in the middle of the network (Figure 1). The ‘dropout’ layer sets a random selection of weights to zero during each training step, which helps prevent overfitting (Srivastava et al., 2014). Our implementation allows users to adjust the shape of the network, but current default settings use 10 dense layers of 256 nodes each with ‘ELU’ activations (Clevert et al., 2015) and a 25% dropout after the fifth layer. We describe performance under varying network width and depth in Figure 5—figure supplement 1. In general, we found that all networks with over four layers perform similarly.

We use the Adam optimizer (Kingma and Ba, 2014) with Euclidean distance as a loss function:(1)loss=(xpredicted-xtrue)2+(ypredicted-ytrue)2.

Individuals are born at a single location, but have inherited their genomes as a mosaic from ancestors spreading geographically into the past (as discussed in, for instance, Wright, 1943; Kelleher et al., 2016; Bradburd and Ralph, 2019). Any signal our method hopes to extract from the data must be due to geographic signal of recent ancestors shared between the test and training datasets. This suggests that any analogous method must quantify, roughly, ‘which modern day populations are most similar to this genome?”. The spatial spread of genetic relatedness both back in time from an individual’s to its ancestors’ locations and forward in time from ancestors to the present-day location of training samples means that even a perfect inference algorithm should have significant uncertainty associated with any predicted location from genetic data, and the magnitude of uncertainty should be in part a function of the dispersal rate of the population. In particular, no such method can infer locations more accurately than the mean dispersal distance, because in most cases an individual’s genome is not informative about where they live relative to their parents. Besides this fundamental limit to uncertainty, error in georeferencing of training samples and in model fitting will introduce additional prediction uncertainty.

We use a windowed analysis across the genome to describe this uncertainty, which is possible thanks to Locator’s computational efficiency. Genealogical relatedness on each contiguous stretch of genome can be described by a sequence of genealogical trees, separated by ancestral recombination events. By running Locator on a particular window of the genome, we restrict inference to a subset of these marginal trees, and hence to a subset of the genetic relationships between test and training samples. Predictions from different regions of the genome can then be visualized as a cloud of points, and the distribution of these points in space gives us a rough idea of the uncertainty associated with an individual-level prediction. Because windowed analyses involve repeated training runs from randomized starting parameters, they also help us to integrate over uncertainty associated with the model fitting process.

Some datasets lack the size or reference alignments necessary to conduct windowed analyses. In this case, we recommend uncertainty be assessed by training replicate models on bootstrapped samples drawn from a single set of unlinked SNPs (that is, resampling SNPs with replacement). Although this procedure does not reduce the number of marginal trees represented in the data, it does allow us to assess uncertainty associated with model training and prediction. In both cases, we summarize uncertainty in predicted locations by estimating a two-dimensional kernel density surface over a set of predicted locations, and provide plotting scripts to visualize the 95%, 50%, and 10% quantiles in geographic space (see Figures 5–7 for examples). The location of an individual can then be predicted as either the location with highest kernel density (the modal prediction) or the geographic center of the cloud of predictions (the mean prediction).

We tested this approach in simulated data and in all empirical datasets. To explore factors affecting the accuracy of predicted locations generated from different regions of the genome, we also examined the relationship between recombination rate and test error from windowed Locator runs on human data from the HGDP panel (Bergström et al., 2020). Recombination rates for each window were estimated by averaging per-base rates from the HapMap project (International HapMap Consortium, 2003).

We first evaluated our method on genotypes from populations simulated by SLiM v3 (Haller and Messer, 2019), using the model of continuous space described in Battey, 2019. We simulated a 50 × 50 unit square landscape with expected density (d) of 5 individuals per unit area, resulting in census sizes of around 12,500. We varied the mean parent-offspring dispersal distance σ across simulations from 0.45 to 3, to create populations with varying levels of isolation by distance. In terms of Wright’s ‘neighborhood size’ (Wright, 1946), defined as Nloc=4⁢π⁢σ2⁢d, this yields populations with neighborhood sizes from 13 to 565. Each diploid individual carried two copies of a 10⁸ bp chromosome on which mutations and recombinations occured at a rate of 10^-8 per bp per generation. Simulations were run until all extant individuals shared a single common ancestor within the simulation at all locations on the genome (i.e., the tree sequence had coalesced). 500 individuals were randomly sampled from the final generation of each simulation for use in model fitting.

We selected 50 individuals from each simulation as a validation set and ran Locator while varying the number of training samples from 10 to 450 and the number of SNPs from 100 to 100,000. The SNPs used were a subset sampled from the full genotype matrix without replacement and thus mimic the semi-random distribution of genome-wide SNPs generated by reduced-representation sequencing approaches like RADseq (Etter et al., 2012). To compare performance with an existing model-based approach, we also ran SPASIBA (Guillot et al., 2016) on the simulation with σ=0.44 using 450 training samples and varying the number of SNPs from 100 to 100,000. Locator was run on a CUDA-enabled GPU and SPASIBA was run on 80 CPU cores. Last, we ran a windowed analysis on the σ=0.63 (neighborhood size ≈25) simulation in Locator using a 2Mbp window size (each window then contains ≈ 8000 SNPs).

We applied Locator to three whole-genome resequencing datasets of geographically widespread samples: (1) 765 mosquitoes from the Anopheles gambiae/coluzzii species complex collected across sub-Saharan Africa (Anopheles gambiae 1000 Genomes Consortium et al., 2017), (2) 5965 samples of the malaria parasite Plasmodium falciparum sequenced from human blood samples collected across Papua New Guinea, southeast Asia, sub-Saharan Africa, and northern South America (Pearson et al., 2019) and (3) whole-genome data for 56 human populations from the Human Genome Diversity Project (Bergström et al., 2020). Genotype calls for the Anopheles dataset are available at https://www.malariagen.net/data/ag1000g-phase1-ar3, for P. falciparum at https://www.malariagen.net/resource/26, and for human data at ftp://ngs.sanger.ac.uk/production/hgdp. We used VCF files as provided with no further postprocessing.

The Plasmodium falciparum dataset is unusual relative to our other empirical examples in that sequencing libraries were prepared from blood samples without filtering for coinfections or isolating individual Plasmodium. Sequence reads returned from short read sequencing then reflect the population of Plasmodium present in a human blood sample, or even multiple lineages of parasite if an individual is co-infected with multiple strains (Zhu et al., 2019), rather than individual Plasmodium. The VCFs we analyzed were prepared by aligning illumina short read sequences to the Plasmodium falciparum reference genome prepared by the Pf3K project (Pf3K Consortium, 2016; https://www.malariagen.net/data/pf3K-5), then calling SNPs in GATK (McKenna et al., 2010). Variant calls then represent the pool of mutations present in the infecting Plasmodium population rather than SNPs in a single individual. We used only field-collected samples from the ‘analysis’ set, as described in Pearson et al., 2019.

For the Anopheles dataset, we ran Locator in 2Mbp windows across the genome with a randomly selected 10% of individuals held out as a test set. We also ran SPASIBA on subsets sampled from the first five million base pairs of chromosome 2L while varying the number of SNPs from 100 to 100,000. For the P. falciparum dataset, we used 500 kb windows and held out 5% of samples from each collection locality as a test set. Last, for humans we used 10Mbp windows and selected three individuals from each HGDP population to hold out as a test set. Window sizes in each case were chosen to include roughly 100,000–200,000 SNPs per window. All empirical analyses were run with default settings (10×256 network size, patience 100, 25% dropout, a random 10% of training samples used for validation).

We also tested Locator’s performance with empirical data when the true location is not represented in the training sample. To do this, we ran a series of models on 10,000 SNPs randomly selected from the first 5Mbp of chromosome 2L in the Anopheles data. For each run, we held out all samples from a given sampling locality from the training set, then predicted the locations of these individuals using the trained model. We also tested this approach while holding out all samples collected in a given country, which eliminates even nearby localities from the training set.

Locator is implemented as a command-line program written in Python: www.github.com/kern-lab/locator. SNP calls for the Anopheles dataset are available at https://www.malariagen.net/data/ag1000g-phase1-ar3, for P. falciparum at https://www.malariagen.net/resource/26, and for the HGDP at ftp://ngs.sanger.ac.uk/production/hgdp. Code to run continuous-space simulations can be found at https://github.com/kern-lab/spaceness/blob/master/slim_recipes/spaceness.slim (Battey, 2019). This publication uses data from the MalariaGEN Plasmodium falciparum Community Project as described in Pearson et al., 2019. Statistical analyses and many plots were produced in R (R Development Core Team, 2018).