The reasonable effectiveness of domain adaptation for inference of introgression

The reasonable effectiveness of domain adaptation for inference of introgression | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results The reasonable effectiveness of domain adaptation for inference of introgression View ORCID Profile Kerry A. Cobb , Megan L. Smith doi: https://doi.org/10.1101/2025.01.17.633659 Kerry A. Cobb 1 Department of Biological Sciences, Mississippi State University , Mississippi State, MS 39762 Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Kerry A. Cobb For correspondence: cobbkerry{at}gmail.com Megan L. Smith 1 Department of Biological Sciences, Mississippi State University , Mississippi State, MS 39762 Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract Supervised machine learning approaches have proven powerful in population genetics. To use such approaches, training data with known inputs and outputs are required. Since such data are generally unavailable in population genetics, researchers typically rely on simulations under the models of interest to train machine learning algorithms. While powerful, this approach depends heavily on the models used to generate training data. Because of the variety and complexity of processes shaping genetic variation, it is inevitable that not all processes important in an empirical system will be included when generating training data. This leads to a mismatch between the data used to train a machine learning algorithm and the data to which the trained model is ultimately applied–i.e., a domain shift– and can negatively impact inference. Here, we train a Convolutional Neural Network (CNN) to detect introgression between sister populations and demonstrate that it has near perfect accuracy when applied to data generated under the models used for training. To evaluate the impacts of domain shifts on inference, we generated new data with introgression from a third, unsampled population into one of the two focal populations (i.e., ghost introgression), and accuracy was substantially reduced on these data. Finally, we used domain adaptation, which aims to train a network that performs well in the presence of a domain shift. Notably, this requires no knowledge of the target or empirical domain. Our domain adaptation network was able to accurately detect introgression, even in the presence of unmodelled ghost introgression. We also applied this approach to empirical data to detect introgression between ABC Island brown bears and other populations of brown bears. Previous work has suggested that introgression between ABC Island bears and polar bears can mislead tests of introgression between populations of brown bears. We found that using domain adaptation reduced support for introgression between geographically isolated populations of brown bears, suggesting that our approach reduces false inferences of introgression due to ghost introgression. I ntroduction Analyses of genomic data from across the tree of life have supported rampant introgression between populations and species ( Mallet et al., 2016 ). Detecting introgression, or the movement of genetic material from one lineage into another, has long interested biologists, as it plays a central role in understanding speciation ( Roux et al., 2016 ), inferring species boundaries ( Jackson et al., 2017 ; Smith and Carstens, 2020 ), developing null models for detecting selection ( Nielsen et al., 2007 ), and more. Given this widespread interest, it is unsurprising that several methods exist for detecting introgression between populations and species. Phylogenetic methods require data from more than two populations, and can only detect introgression between non-sister taxa. Many phylogenetic methods evaluate whether there are deviations from the site patterns expected under models without introgression (e.g., the multispecies coalescent model), and interpret deviations from these expected patterns as evidence of potential introgression (reviewed in Hibbins and Hahn 2022 ). Other phylogenetic methods aim to infer species networks by quantifying the fit of sequence data to various network topologies under a specific model, such as the multispecies coalescent model on networks ( Solís-Lemus and Ané, 2016 ). Detecting introgression between sister taxa, on the other hand, requires population genetic approaches. Population genetic methods for detecting introgression typically fit a model to the data (e.g., Excoffier et al. 2021 ; Gutenkunst et al. 2009 ). Then, migration rates can be inferred and interpreted, or models with and without migration can be compared. All of these methods rely on models. Phylogenetic approaches rely on models either indirectly (i.e., by comparing expectations under a null model to the observed data) or directly (i.e., by assessing the fit of models to the data), while the most popular population genetic approaches do the latter. Models must make assumptions about the processes shaping genetic variation. For example, most methods rely on models that assume an absence of gene flow from un-sampled populations (i.e., ghost introgression, Beerli 2004 ) or that selection has no impact on observed genetic variation. While simplifying assumptions are inevitable in model-based inference, some assumptions are frequently violated in empirical systems, and violations of these assumptions can mislead inference. For example, many empirical studies fail to sample all relevant populations or species either due to extinctions, insufficient funds, difficulty sampling, or a lack of knowledge regarding relevant populations, so ghost introgression is likely common in empirical studies. Furthermore, ghost introgression has been shown to mislead both phylogenetic ( Tricou et al., 2022 ) and population genetic ( Beerli, 2004 ; Strasburg and Rieseberg, 2010 ) approaches for detecting introgression. Similarly, selection appears to influence patterns of variation across the genomes of many species ( Cutter and Payseur, 2013 ; Kern and Hahn, 2018 ) and can mislead population genetic approaches for detecting introgression ( Smith and Hahn, 2024 ). Machine learning approaches offer a powerful alternative to likelihood and Bayesian inference when researchers aim to make inferences from highly heterogenous datasets generated by diverse processes. Specifically, when the models of interest are intractable in likelihood and Bayesian frameworks, supervised machine learning offers a promising solution. Supervised machine learning uses training data—i.e., data with known inputs and outputs—to learn to predict input from output. For example, if a researcher wishes to train a machine learning model to detect introgression, a training dataset consisting of genetic data from pairs of populations or species for which we know whether introgression has occurred is necessary. Of course, such datasets are generally unavailable, especially in the quantities necessary for supervised machine learning (i.e., 1000s of datasets under each model). To circumvent this issue, machine learning approaches in population genetics have typically relied on simulations under the model or models of interest to obtain training data ( Schrider and Kern, 2018 ). This offers a high level of flexibility, as the user can include any processes under which they can simulate data when generating training datasets. Supervised machine learning has seen a wide range of successful applications in population genetics, including detecting introgression (e.g., Schrider et al. 2018a ; Ray et al. 2024 ), inferring population size trajectories (e.g., Sanchez et al. 2021 ), detecting selection (e.g., Hejase et al. 2022 ), and jointly inferring demography and selection (e.g., Sheehan and Song 2016 ). Despite this flexibility, it remains infeasible to simulate every potential process acting in real biological systems. This leads to an inevitable mismatch between the simulated data used to train the machine learning model and the empirical data on which the researcher aims to make predictions. Such a shift is referred to as a domain shift in the machine learning literature and is known to lead to decreased accuracy. In population genetics and phylogenetics, a few studies have evaluated the impacts of domain shifts on inference with mixed results. For example, Sheehan and Song (2016) found that applying a network to datasets generated under higher rates of recombination than used in training led to decreased accuracy when detecting selection but did not substantially impact inferences of population sizes. On the other hand, applying a network to datasets generated under more extreme bottlenecks negatively impacted population size inference but had minimal impacts when detecting selection. In another study, Schrider et al. (2018a) found minimal impacts of a mis-specified demographic model on attempts to detect introgression. Thus, it remains unclear the extent to which machine learning approaches in population genetics are susceptible to domain shifts. While domain shifts are potentially a major issue in supervised machine learning, there are approaches that aim to minimize the impacts of such shifts on inference. Domain adaptation approaches aim to construct classifiers that perform well even in the presence of a shift between the source and target domains (reviewed in Wilson and Cook 2020 ). One of the most popular approaches is domain-invariant feature extraction, which aims to minimize the differences in features extracted from the source domain (i.e., training domain) and the target domain (i.e., empirical domain) This is accomplished by extracting features that perform well at classification in the source domain, but cannot be used to distinguish between the source and target domains (e.g., Ganin et al. 2016 ). Notably, this approach does not require any labeled data (i.e., data with a known history) from the target domain and is therefore potentially powerful in population genetic settings. Mo and Siepel (2023) demonstrated that both background selection and mis-specification of the demographic model mislead machine learning approaches for classifying selective sweeps and estimating population recombination rates. However, by using domain-invariant feature extraction, they were able to arrive at accurate inferences in the presence of these processes, even though they were not modeled in the training data. This suggests that domain adaptation may be a promising path forward in population genetic inference. Here, we first construct a Convolutional Neural Network (CNN) to detect introgression between sister populations. We then evaluate the impacts of a common model violation–ghost introgression–on inferences using this network. Next, we apply a domain adaptation approach to build a new network that performs well in the presence of ghost introgression. Finally, we apply this network to empirical data from brown bears and demonstrate that it reduces the rate of likely false positives. M ethods General Approach In a general supervised machine learning setting, labeled training data (i.e., source data ) are used to train a machine learning algorithm. In the case of detecting introgression, labels indicate whether or not introgression has occurred between the focal populations. These data are used as input into a machine learning algorithm, which extracts meaningful features and uses these features to classify data as having been collected from populations that have been completely isolated or have experienced introgression. After a network is trained, it can be applied to target data . Target data are usually empirical data without known labels, and the goal is to predict a label for the target data. When the models used to generate source data do not capture the processes shaping variation in the target data, a domain shift has occurred. Domain adaptation techniques aim to train classifiers that work well in the target domain in the presence of a domain shift between the source and target domains ( Wilson and Cook, 2020 ). Here, we used adversarial domain adaptation, specifically Conditional Adversarial Domain Adaptation (CDAN, Long et al. 2018 ). Briefly, a traditional Convolutional Neural Network (CNN) performing a classification task has two primary layers: a feature extractor that extracts features from the input data and a classifier that learns to distinguish among classes using these features. Adversarial approaches have an additional layer—a discriminator that aims to distinguish between source and target domains ( Fig. 1 ). While the discriminator attempts to learn to distinguish domains, the feature extractor tries to learn features that cannot be used to distinguish among domains. By extracting such invariant features, it should be possible to build a classifier that can make accurate predictions in both the source and target domains without need for any labeled data from the target domain. CDAN builds on the Deep Adaptation Network (DANN, Ganin et al. 2016 ) by using a domain discriminator that is conditional on the current predictions of the label classifier and captures cross-covariance between feature representations and classifier predictions. This overcomes issues of multimodality in the data and has been found to improve performance relative to DANN ( Long et al., 2018 ). Below, we describe our approach for simulating training and target data, the details of our network, and an empirical test using data from brown bears. Download figure Open in new tab Figure 1. Overview of our CDAN architecture. Genetic data from a pair of populations are summarized as joint Site Frequency Spectra (SFS) which are used as input into the network. The feature extractor extracts features from the SFS. The classifier uses these features to detect introgression, while the discriminator uses these features features along with the classifier predictions to distinguish the source and target domains. However, the gradient reversal layer (GRL) inverts the loss function during back propagation, which discourages the feature extractor from learning features that are not shared by both domains, minimizing the domain discrimination. Simulations To generate data for training and testing our machine learning approach for detecting introgression, we used msprime ( Baumdicker et al., 2022 ). For our source data, we focused on two demographic models: a divergenceonly model and a model with gene flow upon secondary contact between the two focal populations ( Fig. 2 ). We simulated a 500,000 bp region of a single chromosome with a mutation rate µ and a recombination rate r . The populations diverged at time τ and shared a single constant effective population size. For the secondary contact model, migration began halfway between the time of divergence and the present and continued until the present. Population sizes, divergence times ( τ ), and migration rates were drawn from uniform priors ( Table 1 ). We sampled 10 diploid individuals from each population at the present and summarized the data as folded joint site frequency spectra (SFS). We then normalized each SFS output from msprime such that all values sum to one. View this table: View inline View popup Download powerpoint Table 1. Parameter values and ranges of uniform probability distributions for simulations in msprime . All times are in units of generations. Mutation and recombination rates are in units of per site per generation. The migration rate is the effective population size N e times the expected number of migrants per generation m . The pulse migration proportion is the proportion of lineages that move from one population into another at time τ pulse looking backward in time. Download figure Open in new tab Figure 2. Models used to simulate training and testing data. In the source domain, we model two taxa with or without migration upon secondary contact. In the target domain, we expand upon these models by adding introgression into taxon 2 from unsampled taxon G –a ghost lineage. The effective population size N e is shared among all populations. Divergence between taxa 1 and 2 occurs at time τ , and divergence between the ancestor of these taxa and taxon G occurs at time τ ghost . When migration occurs, it begins at time τ migration and continues until the present at a rate m . Introgression from the ghost population is modeled as a pulse event in which a proportion γ of lineages move from taxon 2 to taxon G (backwards in time) at time τ pulse . Next, we simulated data under an alternative model which we treat as the target domain. Data from the target domain are typically unlabeled. However it is useful to have target data with known labels to evaluate the impact of mis-specification on a standard machine learning approach and the effectiveness of domain adaptation. While known, these labels are not used for inference. Specifically, we simulated data under a model with introgression from an un-sampled “ghost” population into one of the two focal populations. Our ghost introgression models were identical to the two models described above, except they also included a third unsampled population that diverged from the two focal populations at time τ ghost . Looking backward in time, at time τ pulse , a proportion of lineages γ from one of the focal populations moved into the ghost lineage ( Fig. 2 ). The γ, τ ghost , and τ pulse parameters were drawn from uniform priors ( Table 1 ). All other aspects of the models were parameterized as in the source models. Together with the source model parameters, we refer to these parameters as the “general scenario”. As above, we summarized the simulated data as a normalized, folded joint SFS for the two focal populations with the recipient of ghost introgression on the row axis. Network We used these simulated data as input for training and testing a CDAN network. A CDAN has three components: a feature extractor which extracts features from the data, a classifier which learns to classify data using the encoded features, and a discriminator which attempts to distinguish between datasets from the source and target domains ( Fig. 1 ). We implemented our CDAN using the python package adapt ( de Mathelin et al., 2021 ), along with functions from tensorflow ( Abadi et al., 2015 ) using the keras API ( Chollet et al., 2015 ). For the feature extractor, we used two successive blocks containing a 2D convolutional layer with a 3×3 filter, a 1×1 stride, and ReLU activation followed by 2D max pooling with a 2×2 pool size and a 2×2 stride. These blocks were followed by flattening and a dense layer with 20 neurons and ReLU activation. For the classifier, we used a dense layer with 100 neurons and ReLU activation followed by a dense layer with 2 neurons and softmax activation. For the discriminator, we used a dense layer with 100 neurons followed by another with 50 neurons, each with ReLU activation and, finally, a dense layer with a single neuron with a sigmoid activation function. We used a categorical cross-entropy loss function for the classifier. We used the Adam optimizer with a learning rate of 0.001 for each component of the network. The CDAN parameter λ is a trade-off parameter that controls the relative weight given to correctly classifying the source data versus minimizing differences between the source and target features. We used a variable λ which was initially set to zero and was then increased linearly up to its specified value over the first 1000 mini-batches across epochs. Training and Evaluation We used data simulated under the models described above to train and evaluate our networks. Under the source domain (i.e., with no ghost introgression), we simulated 20,0000 training datasets, 1,000 test datasets, and 1,000 validation datasets. Under the target domain (i.e., with ghost introgression), we simulated simulated 100 training datasets and 1,000 test datasets. In order to compare networks trained without domain adaptation to networks trained with domain adaptation, we varied the value of the trade-off parameter λ. When λ is set to zero, it is equivalent to using only the feature extractor and classifier, thus having no domain adaptation. We trained 10 replicate networks with λ set to zero and evaluated the performance of these networks on test data generated under both the source and target domain. Then, we trained 10 replicate networks with a maximum λ of 1 to evaluate the effectiveness of domain adaptation and evaluated these networks on the same test datasets. All networks were trained for 50 epochs using a batch size of 16. Network performance was evaluated using confusion matrices and receiver operator curves (ROC). Additionally, we performed principal components analysis (PCA) of feature embeddings for all test datasets. A feature embedding is a vector of learned weights output from the feature extractor which encodes the signal contained in the input data. Visualizing these embeddings allows us to assess differences in the features extracted from the source and target domains. If a domain shift has occurred, we expect to see differences between these distributions. However, if domain adaptation is successful, such differences should be diminished. Empirical Application To Brown Bears Brown bears are widely distributed across the Northern hemisphere and exhibit a high degree of population structure ( de Jong et al., 2023 ). Beginning around the time of the last glacial maximum, several distinct populations diverged ( de Jong et al., 2023 ). Some of these populations have experienced introgression from polar bears, with the most well-known case being brown bears from the Admiralty, Baranof, and Chichagof (ABC) islands in Southeastern Alaska ( Cahill et al., 2015 ; Kumar et al., 2017 ; Wang et al., 2022 ; Lan et al., 2022 ). Introgression between adjacent but divergent brown bear populations is plausible, whereas introgression between non-adjacent populations is unlikely. However, introgression has been inferred between very distant populations when ghost introgression is not considered, making brown bears a good test case for our approach ( Tricou et al., 2022 ). We simulated source and target data under a scenario matching the evolutionary history of polar bears and brown bears and trained models using the same network architecture as above in order to evaluate the performance of our approach in this context. We then applied our approach to test for introgression between ABC Island bears and seven other brown bear populations using networks trained with and without domain adaptation. In this case, our target data are unlabeled empirical data. To generate simulated data for training, validation, and testing under the “bear scenario”, we used the same simulation procedure as above, changing only the parameters of the models ( Table 1 ). In addition to simulations for training our classifier in the source domain, we simulated data meant to mimic the target domain to evaluate the efficacy of our approach in this setting. In this case, in the target domain we anticipate introgression (forward in time) from the polar bear lineage into the ABC Island bears. Polar bears are estimated to have diverged from brown bears 0.38–1.05 mya. ( Kutschera et al., 2014 ; de Jong et al., 2023 ; Abella et al., 2012 ; Zou et al., 2022 ). This is equivalent to 30,400–84,000 generations assuming a mean generation time of 12.5 years as estimated by Wang et al. (2022) . The divergence times between pairs of brown bear populations have been estimated to be between 25,000–80,000 years or 2,000–6,400 generations ago assuming the same mean generation time ( de Jong et al., 2023 ). We drew effective population sizes from a uniform prior spanning 1,000–20,000, which encompasses values used in previous simulation studies ( de Jong et al., 2023 ). We used a mutation rate of 1 e − 8 which approximates the estimate of 0.84e-8 by Wang et al. (2022) . We followed de Jong et al. (2023) in setting the recombination rate to 1 e − 8. As before, we simulated 20,000 training datasets, 1,000 test datasets and 1,000 validation datasets under the source domain (i.e., with no polar bear introgression), and we simulated 100 training datasets and 1,000 test datasets under the target domain (i.e., with polar bear introgression). We summarized these simulated data as normalized, folded joint SFS. We trained and tested networks using the same procedure used for the general scenario with one minor modification. When training networks with domain adaptation, we used a maximum λ of 0.5 because during preliminary analyses we observed that a λ of 0.5 resulted in better, more stable training performance relative to a λ of 1.0 for the bear scenario. We trained 10 replicate networks for each λ value (i.e., λ=0 for no domain adaptation and λ=0.5 for domain adaptation). We evaluated training performance as we did for the general scenario. In order to train domain adapted networks with empirical target data and test for introgression between ABC island bears and the seven other brown bear populations, we used publicly available whole genome re-sequencing data ( Fig. 3 ) ( de Jong et al., 2023 ; Benazzo et al., 2017 ; Barlow et al., 2018 ; Cahill et al., 2013 , 2015 ; Liu et al., 2014 ; Miller et al., 2012 ; Taylor et al., 2018 ). For each of the eight populations, we arbitrarily selected 10 samples and obtained all of the available Illumina sequencing run data for each sample from the European Nucleotide Archive (see Table S2 for accession IDs). We used fastp to remove any remaining adapter sequences and to remove low quality sequences using the default parameters ( Chen, 2023 ). We mapped the filtered and trimmed reads to the 36 largest autosomal scaffolds (assembled chromosomes) of the brown bear genome published by ( Armstrong et al., 2022 ) using BWA-MEM2 ( Li, 2013 ). We merged any mapped reads from samples that were split across different runs using SAMtools ( Danecek et al., 2021 ). To call variants from the mapped reads we used BCFtools , running the mpilup command with a max depth of 250 and without INDEL calling ( Danecek et al., 2021 ). The Nextflow ( Di Tommaso et al., 2017 ) scripts used for processing the bear data are archived at https://doi.org/10.5281/zenodo.14629180 . For each of the seven population pairs, we filtered any sites that had missing data and retained only bi-allelic sites among the samples within the two populations using BCFtools . We then summarized the filtered data from each chromosome as a folded, joint SFS with ABC Island bears on the row axis using δ a δ i ( Gutenkunst et al., 2009 ). When the classification of an allele as the major or minor allele in one population versus the other is ambiguous, δ a δ i adds 0.5 to each of the two possible cells in the folded, joint SFS. This differs from msprime where alleles are always classified as the minor allele in the population on the row axis if classification is ambiguous. To make the SFS output by δ a δ i consistent with the SFS output by msprime , we summed both counts of ambiguously classified alleles, replaced the minor allele count in the population on the row axis with this sum and the major allele count with 0. We then normalized the allele counts as above. For each population pair, we trained 10 replicate networks with a λ of zero and 10 replicate networks with a λ of 0.5 using the same network architecture and procedure as above. For the source training data we re-used the 20,000 training and 1,000 validation data sets used above under the bear scenario. For the target training data we used the 36 SFSs computed from each chromosome. We then computed softmax probabilities for each chromosome in each population pair independently using the trained network. For classifying a chromosome in a population pair as having introgression or not, we applied classification thresholds of 0.5, 0.6, 0.7, 0.8, and 0.9 to the mean softmax probabilities across the ten replicate networks. Download figure Open in new tab Figure 3. Localities and population assignments for brown bear samples used in this study. R esults General Scenario Simulation Both migration and ghost introgression leave clear signals in the simulated datasets. Datasets generated under the source domain had an average of 8,917 Single Nucleotide Polymorphisms (SNPs), and migration between populations led to increased allele sharing in the SFS, whereas, in the absence of migration, there were more alleles unique to each population ( Fig. 1 ). Datasets generated under the target domain (i.e., with ghost introgression) contained more SNPs on average (22,032). Furthermore, ghost introgression tended to skew the SFS, resulting in more low-frequency alleles unique to the recipient population ( Fig. 1 ). Our CNN without domain adaptation (i.e., λ = 0) had high power and a near-zero false positive rate in the absence of a domain shift. Introgression was never inferred for datasets lacking introgression, and networks failed to detect introgression when it was present in a mean of 0.14% of datasets across replicate networks (Fig. S1a, Table S1). The mean AUC score was greater than 0.999, indicating a near perfect classifier ( Fig. 4b ). Download figure Open in new tab Figure 4. Results under the general simulation scenario on datasets generated under the target domain (i.e., with ghost introgression). Confusion matrices show the proportion of test replicates in each category without (a) and with (d) domain adaptation. Values were averaged across the ten replicate networks. (b) ROC curves for the source (blue) and target (orange) test datasets without (b) and with (e) domain adaptation. The average AUC across all replicates is reported. (c) PCA of feature embeddings from the source (blue) and target (orange) test datasets without (c) and with (f) domain adaptation. All values are rounded to the nearest hundredths place. See Table S1 for original values. However, in the presence of ghost introgression, the network trained without domain adaptation performed substantially worse. Introgression was incorrectly inferred when it was absent in a mean of 28.10% of datasets, and networks failed to infer introgression when it was present in a mean of 17.98% of datasets ( Fig. 4a ). The mean AUC score was 0.86 with a high degree of variation observed in the ROC curves ( Fig. 4b ). Furthermore, the PCA of source and target embeddings indicated a high degree of mismatch between the embedded features of the two domains ( Fig. 4c ). Next, we evaluated the impact of using domain adaptation on inferences with and without simulation misspecification. In the absence of simulation misspecification, using domain adaptation had no negative impact on network performance. There was only a single false positive across all replicates, and networks failed to detect introgression in a mean of 0.3% of datasets (Fig. S1b). The mean AUC score across replicates was greater than 0.999, again indicating a near perfect classifier ( Fig. 4e ). In the presence of ghost introgression, using domain adaptation substantially improved accuracy compared to the networks without domain adaptation. Introgression was incorrectly inferred when it was absent in a mean of only 0.86% of datasets across replicates, and networks failed to infer introgression when it was present in a mean of only 5.04% of datasets across replicates ( Fig. 4d ). The mean AUC score was 0.994, indicating a near perfect classifier even in the presence of simulation misspecification ( Fig. 4e ). The PCA of source and target embeddings also indicated a much greater degree of overlap between the domains than was seen with the network trained without domain adaptation ( Fig. 4f ). Classification accuracy increased early in training, and the accuracy remained high throughout (Fig. S2). On the other hand, discriminator accuracy often began high, but decreased quickly, fluctuating around 0.5 for most of training. This indicates that our network performs as desired, extracting features that facilitate classification while minimizing differences between the source and target domains. Bear Scenario Simulation The bear scenario presents a more challenging problem than the general scenario due to more recent divergence times, so we also evaluated the performance of networks with and without domain adaptation under this scenario. The source datasets had a mean of 21,851 SNPs, while target datasets had a mean of 34,112 SNPs. Unsurprisingly, even in the absence of simulation misspecification, networks performed slightly worse in this scenario, but error rates remained low. Networks trained without domain adaptation (i.e., λ = 0) falsely inferred introgression when it was absent in a mean of 4.14% of datasets and failed to detect introgression when it was present in a mean of 6.22% of datasets (Fig. S1c). The mean AUC score for datasets without simulation misspecification was 0.991, indicating a highly accurate classifier ( Fig. 5b ). Download figure Open in new tab Figure 5. Results under the bear simulation scenario on datasets generated under the target domain (i.e., with ghost introgression). Confusion matrices show the proportion of test replicates in each category without (a) and with (d) domain adaptation. Values were averaged across the ten replicate networks. (b) ROC curves for the source (blue) and target (orange) test datasets without (b) and with (e) domain adaptation. The average AUC across all replicates is reported. (c) PCA of feature embeddings from the source (blue) and target (orange) test datasets without (c) and with (f) domain adaptation. All values are rounded to the nearest hundredths place. See Table S1 for original values. Network performance was again substantially worse in the presence of ghost introgression. Introgression was incorrectly inferred when it was absent in a mean of 17.6% of datasets across replicates, and networks failed to infer introgression when it was present in a mean of 2.08% of datasets ( Fig. 5a ). However, the mean AUC score across replicates for tests with ghost introgression was 0.98, which is nearly equal to the mean AUC score for tests using source data, likely due to high performance on the positive class ( Fig. 5b ). The PCA of source and target domain embeddings indicated some degree of mismatch between the two domains ( Fig. 5c ). As in the general scenario, domain adaptation did not negatively impact performance in the absence of simulation misspecification. Networks falsely inferred introgression when it was absent in a mean of 4.7% of datasets and failed to detect introgression when it was present in a mean of 5.68% of datasets (Fig. S1d). The mean AUC score across replicates for datasets without simulation misspecification was 0.99 ( Fig. 5e ). As in the general scenario, in the presence of ghost introgression, networks with domain adaptation sub-stantially outperformed networks without domain adaptation. The trained networks misclassified a mean of 4.38% of datasets simulated without introgression and 4.82% of datasets simulated with introgression ( Fig. 5d ). The mean AUC score across replicates was 0.989 ( Fig. 5e ). The PCA of source and target domain embeddings had a slightly greater degree of overlap than the embeddings from the network without domain adaptation ( Fig. 5f ). Empirical Bear Data Next, we applied networks with and without domain adaptation to data from the ABC Island bears and seven other brown bear populations, with the expectation that ghost introgression from polar bears to the ABC Island bears could mislead inferences. After mapping the bear sequence data to the reference and filtering, the mean number of SNPs per chromosome in each population pair tested for introgression was 497,657.4. Without domain adaptation, the softmax probability of introgression across replicates was high across most chromosomes and population pairs, with a mean of 0.76 (SD = 0.30) ( Fig. 6a ). Notably, inferred probabilities of introgression were high even between the ABC Island bears and geographically distant populations from Asia and Europe. When applying a classification threshold of 0.5, most chromosomes from most populations are classified as having introgression (Fig. S3a). Even at with a classification threshold of 0.9, the most conservative threshold that we applied, a large number of chromosomes in geographically distant populations were classified as having introgression ( Fig. 6c ). Download figure Open in new tab Figure 6. Results from tests of introgresssion between the ABC Island bears and other brown bear populations. Rows indicate the brown bear population, and columns correspond to the 36 chromosomes tested. (a-b) Average softmax probabilities across replicate networks without (a) and with (b) domain adaptation. (c-d) A 90% threshold was applied to the softmax probabilities, such that a chromosome was only classified as introgressed if the average softmax probability of introgression across all replicates was > 0.9. Based on this threshold, chromosomes are classified as introgressed (yellow) or not introgressed (blue). Results are shown without (c) and with (d) domain adaptation. Training networks with domain adaptation reduced support for introgression between geographically isolated populations of brown bears. The mean softmax probabilities of introgression across replicates were much lower for most chromosomes in all populations, with a mean of 0.49 (SD = 0.44) ( Fig. 6b ). Further-more, the number of datasets classified as having introgression using a classification threshold of 0.5 was much lower (Fig. S3b). When applying the most conservative classification threshold of 0.9, the number of positive tests is greatly reduced. Notably, there are fewer tests supporting introgression between the ABC Island bears and geographically isolated populations ( Fig. 6d ). Introgression is inferred between the ABC Island bears and other bear populations from North America with greater frequency. D iscussion Machine learning approaches used in population genetics typically rely on simulated training data due to the absence of real labeled datasets. As a consequence, the processes generating empirical data may not be modeled, and the simulated data used to train the machine learning model may differ substantially from empirical data. A mismatch between the training data distribution, referred to as the source domain, and the empirical data distribution, referred to as the target domain, can reduce the accuracy of inferences made with these tools ( Wilson and Cook, 2020 ). We demonstrated the effectiveness of domain adaptation in overcoming this issue in inferences of introgression from population genetic data. Specifically, we tested the performance of CDAN, a type of adversarial domain adaption ( Long et al., 2018 ). With this approach, the network is trained to extract features shared across the source and target domains that can classify the data into the desired categories. In our case, the aim was to find features that could be used to classify the data as having past introgression between populations or not, even in the presence of unmodeled ghost introgression, a known cause of error in tests for introgression ( Tricou et al., 2022 ; Beerli, 2004 ; Strasburg and Rieseberg, 2010 ). We found that unmodeled ghost introgression negatively impacted inference using a standard CNN. However, using CDAN, we could detect introgression with very high accuracy, even in the presence of ghost introgression ( Fig. 4 ). The CDAN was able to achieve this high level of accuracy using a relatively small amount of unlabeled data from the target domain, which bodes well for real-world applications of this approach. The much higher level of accuracy achieved with the CDAN indicates that it was able to effectively learn features informative for classification of the data that are shared between the source and target domains. Additionally, the features extracted by the CDAN exhibit a higher degree of similarity between the source and target domains compared to features extracted by a traditional CNN, demonstrating the alignment of the learned features extracted during training ( Fig. 4c,f ). Consistent with the findings of Mo and Siepel (2023) , our CDAN performed as well as the traditional CNN in the absence of a domain shift, suggesting that there is no cost in terms of accuracy to performing domain adaptation. When applied to empirical data from brown bears, the network trained without domain adaptation produced highly implausible results. We do not expect for introgression to have occurred between the ABC Island bears and bears from Asia and Europe. The Bering land bridge that permitted the migration of brown bears into North America was inundated by rising sea levels by approximately 11,000 years ago, isolating North American brown bears from other brown bear populations ( de Jong et al., 2023 ). The occurrence of gene flow between ABC Island brown bears and mainland brown bears in North America, on the other hand, has been previously demonstrated ( Cahill et al., 2013 ). Despite this expectation, the original network inferred introgression between Asian and European brown bears and the ABC island bears across many chromosomes ( Fig. 6 ). Using domain adaptation reduced support for introgression, particularly between geographically isolated populations ( Fig. 6 ), and we suspect that these results are far more reliable than those without domain adaptation. In general, we expect that domain adaptation should work well when there are features that are both informative for classification and shared across domains. Introgression has a clear and distinctive impact on the joint SFS, leading to increased allele sharing between populations, while ghost introgression skews the SFS, resulting in more low-frequency alleles unique to the recipient population ( Fig. 1 ). Since ghost introgression clearly alters the SFS, but does not produce patterns that exactly match those produced by introgression between the focal populations, it is relatively straightforward for the network to learn features useful for classification but shared across domains. Applying domain adaptation would likely be less effective if the patterns produced by ghost introgression more closely matched those produced by introgression between the focal populations. Visualizing the SFS across correctly and incorrectly classified datasets clearly indicates that this is not the case here (Fig. S4). Several questions remain, particularly concerning the application of our CDAN to empirical data. Below, we discuss classification thresholds, additional sources of misspecification, and the use of summary statistics in more detail. When applying our CDAN network to simulated data, the error rate at a classification threshold of 0.5 was very low. However, in the empirical data, using a threshold of 0.5 produced implausible results. Therefore, we also presented results using more conservative classification thresholds and found the most conservative threshold to produce more plausible results. While similar thresholds have been applied in other machine learning approaches for detecting introgression ( Schrider et al., 2018b ), selecting an appropriate threshold is not straightforward, and it is challenging to justify the use of any particular threshold. If we consider the scenario without introgression to be the null hypothesis, it is reasonable to require a higher burden of proof to reject it. It is important to keep in mind that softmax values do not represent true probabilities and are known to be poor indicators of confidence level. They can be even more unreliable indicators when applied outside of the source domain ( Guo et al., 2017 ; Pearce et al., 2021 ). Until these patterns are better understood or better solutions exist, researchers should exercise caution when interpreting softmax values. In our simulations, there was a single source of misspecification–ghost introgression. However, myriad other sources of misspecification may impact empirical datasets. In addition to ghost introgression, background selection has been shown to negatively impact tests for introgression ( Smith and Hahn, 2024 ). Additional sources of misspecification could include genotyping error, ghost introgression from multiple populations, population size changes, and more. Fortunately, our domain adaptation approach should be agnostic to the source of misspecification and should work as long as features exist that are shared across source and target domains which can be used for classification. However, there may be sources of model misspecification that alter the data in ways that domain adaptation cannot overcome. We also simulated a fairly modest level of misspecification, but it has been shown that domain adaptation is sensitive to the degree of misspecification ( Mo and Siepel, 2023 ). Future work should explore different sources of model misspecification, interactions between sources of model misspecification, and the limits of domain adaptation in accommodating model misspecification. Here, we used a summary statistic, the site frequency spectrum, as input into the network. In our tests using simulated data, this resulted in near-perfect accuracy. However, summary statistics cannot capture all of the signal of evolutionary history that may be present in genetic data. One of the exciting possibilities of using CNNs for analyzing genetic data is that the raw DNA alignment can be used as input, which could in theory result in less loss of information as compared to the use of summary statistics ( Flagel et al., 2019 ). Analyzing alignments directly could also facilitate domain adaptation by increasing the variety of features that could be learned for classification, which could in turn broaden the number of features available for classification and shared across domains. However, CNNs using raw sequence alignments are more challenging to implement as they come with greater computational burden, and trained models cannot be as readily reused for datasets with different alignment lengths. Using combinations of several different summary statistics as input may yield improvements in domain adaptation while remaining flexible and easier to implement ( Schrider et al., 2018b ), and future work should explore these options. C onclusions Our results demonstrate both the fragility of traditional machine learning approaches in the presence of domain shifts and the effectiveness of domain adaptation in population genetics. The high error rates associated with a relatively mild (and likely common) model misspecification warrant caution when applying supervised machine learning approaches to population genetic data. Avoiding model misspecification is impossible, as all models are necessarily simplifications of reality, and our results suggest that standard supervised learning approaches may not be robust to common model violations. Despite these concerning results, our results also suggest a promising path forward. The domain adaptation approach applied here is computationally tractable, requires no labeled data from the target domain, and apparently does not decrease accuracy even in the absence of misspecification. We are cautiously optimistic, and suggest that future machine learning approaches in population genetics should strongly consider the use of unsupervised domain adaptation. C ode A vailability All code used for this project is available at: https://doi.org/10.5281/zenodo.14681223 A cknowledgements Thank you to members of the Smith Lab for providing feedback on this manuscript. R eferences ↵ Abadi , M. , A. Agarwal , P. Barham , E. Brevdo , Z. Chen , C. Citro , G. S. Corrado , A. Davis , J. Dean , M. Devin , S. Ghemawat , I. Goodfellow , A. Harp , G. Irving , M. Isard , Y. Jia , R. Jozefowicz , L. Kaiser , M. Kudlur , J. Levenberg , D. Mané , R. Monga , S. Moore , D. Murray , C. Olah , M. Schuster , J. Shlens , B. Steiner , I. Sutskever , K. Talwar , P. Tucker , V. Vanhoucke , V. Va-sudevan , F. Viégas , O. Vinyals , P. Warden , M. Wattenberg , M. Wicke , Y. Yu , and X. Zheng . 2015 . Tensorflow: large-scale machine learning on heterogeneous systems . Software available from tensorflow.org . ↵ Abella , J. , D. Alba , J. Robles Gimenez , A. Valenciano Vaquero , C. Rotgers , R. Carmona , P. Montoya , and J. Morales . 2012 . Kretzoiarctos gen. nov., the oldest member of the giant panda clade . PLOS One 7 : e48985 . OpenUrl CrossRef PubMed ↵ Armstrong , E. E. , B. W. Perry , Y. Huang , K. V. Garimella , H. T. Jansen , C. T. Robbins , N. R. Tucker , and J. L. Kelley . 2022 . A beary good genome: haplotype-resolved, chromosome-level assembly of the brown bear (Ursus arctos) . Genome Biology and Evolution 14 : evac125 . OpenUrl CrossRef PubMed ↵ Barlow , A. , J. A. Cahill , S. Hartmann , C. Theunert , G. Xenikoudakis , G. G. Fortes , J. L. A. Paijmans , G. Rabeder , C. Frischauf , A. Grandal-d’Anglade , A. García-Vázquez , M. Murtskhvaladze , U. Saarma , P. Anijalg , T. Skrbinšek , G. Bertorelle , B. Gas-parian , G. Bar-Oz , R. Pinhasi , M. Slatkin , L. Dalén , B. Shapiro , and M. Hofreiter . 2018 . Partial genomic survival of cave bears in living brown bears . Nature Ecology & Evolution 2 : 1563 – 1570 . OpenUrl CrossRef PubMed ↵ Baumdicker , F. , G. Bisschop , D. Goldstein , G. Gower , A. P. Ragsdale , G. Tsambos , S. Zhu , B. Eldon , E. C. Ellerman , J. G. Galloway , et al. 2022 . Efficient ancestry and mutation simulation with msprime 1.0 . Genetics 220 : iyab229 . OpenUrl CrossRef PubMed ↵ Beerli , P. 2004 . Effect of unsampled populations on the estimation of population sizes and migration rates between sampled populations . Molecular Ecology 13 : 827 – 836 . OpenUrl CrossRef PubMed Web of Science ↵ Benazzo , A. , E. Trucchi , J. A. Cahill , P. M. Delser , S. Mona , M. Fumagalli , L. Bunnefeld , L. Cornetti , S. Ghirotto , M. Girardi , L. Ometto , A. Panziera , O. Rota-Stabelli , E. Zanetti , A. Karamanlidis , C. Groff , L. Paule , L. Gentile , C. Vilà , S. Vicario , L. Boitani , L. Orlando , S. Fuselli , C. Vernesi , B. Shapiro , P. Ciucci , and G. Bertorelle . 2017 . Survival and divergence in a small group: The extraordinary genomic history of the endangered Apennine brown bear stragglers . Proceedings of the National Academy of Sciences 114 : E9589 – E9597 . OpenUrl Abstract / FREE Full Text ↵ Cahill , J. A. , R. E. Green , T. L. Fulton , M. Stiller , F. Jay , N. Ovsyanikov , R. Salamzade , J. S. John , I. Stirling , M. Slatkin , and B. Shapiro . 2013 . Genomic evidence for island population conversion resolves conflicting theories of polar bear evolution . PLOS Genetics 9 : e1003345 . OpenUrl CrossRef ↵ Cahill , J. A. , I. Stirling , L. Kistler , R. Salamzade , E. Ersmark , T. L. Fulton , M. Stiller , R. E. Green , and B. Shapiro . 2015 . Genomic evidence of geographically widespread effect of gene flow from polar bears into brown bears . Molecular Ecology 24 : 1205 – 1217 . OpenUrl CrossRef ↵ Chen , S. 2023 . Ultrafast one-pass fastq data preprocessing, quality control, and deduplication using fastp . iMeta 2 : e107 . OpenUrl CrossRef ↵ Chollet , F. et al. 2015 . Keras . ↵ Cutter , A. D. and B. A. Payseur . 2013 . Genomic signatures of selection at linked sites: unifying the disparity among species . Nature Reviews Genetics 14 : 262 – 274 . OpenUrl CrossRef PubMed ↵ Danecek , P. , J. K. Bonfield , J. Liddle , J. Marshall , V. Ohan , M. O. Pollard , A. Whitwham , T. Keane , S. A. McCarthy , R. M. Davies , and H. Li . 2021 . Twelve years of samtools and BCFtools . GigaScience 10 : giab008 . OpenUrl CrossRef PubMed ↵ de Jong , M. J. , A. Niamir , M. Wolf , A. C. Kitchener , N. Lecomte , I. V. Seryodkin , S. R. Fain , S. B. Hagen , U. Saarma , and A. Janke . 2023 . Range-wide whole-genome resequencing of the brown bear reveals drivers of intraspecies divergence . Communications Biology 6 : 1 – 16 . OpenUrl CrossRef PubMed ↵ de Mathelin , A. , F. Deheeger , G. Richard , M. Mougeot , and N. Vayatis . 2021 . ADAPT: awesome domain adaptation python toolbox . arXiv 2107.03049. ↵ Di Tommaso , P. , M. Chatzou , E. W. Floden , P. P. Barja , E. Palumbo , and C. Notredame . 2017 . Nextflow enables reproducible computational workflows . Nature Biotechnology 35 : 316 – 319 . OpenUrl CrossRef PubMed ↵ Excoffier , L. , N. Marchi , D. A. Marques , R. Matthey-Doret , A. Gouy , and V. C. Sousa . 2021 . Fastsimcoal2: demographic inference under complex evolutionary scenarios . Bioinformatics 37 : 4882 – 4885 . OpenUrl CrossRef PubMed ↵ Flagel , L. , Y. Brandvain , and D. R. Schrider . 2019 . The unreasonable effectiveness of convolutional neural networks in population genetic inference . Molecular Biology and Evolution 36 : 220 – 238 . OpenUrl CrossRef PubMed ↵ Ganin , Y. , E. Ustinova , H. Ajakan , P. Germain , H. Larochelle , F. Laviolette , M. March , and V. Lempitsky . 2016 . Domain-adversarial training of neural networks . Journal of Machine Learning Research 17 : 1 – 35 . OpenUrl ↵ D. Precup and Y. W. Teh , eds Guo , C. , G. Pleiss , Y. Sun , and K. Q. Weinberger . 2017 . On calibration of modern neural networks . Pages 1321 – 1330 in Proceedings of the 34th International Conference on Machine Learning ( D. Precup and Y. W. Teh , eds.) vol. 70 of Proceedings of Machine Learning Research PMLR. OpenUrl ↵ Gutenkunst , R. N. , R. D. Hernandez , S. H. Williamson , and C. D. Bustamante . 2009 . Inferring the joint demographic history of multiple populations from multidimensional snp frequency data . PLOS Genetics 5 : e1000695 . OpenUrl CrossRef ↵ Hejase , H. A. , Z. Mo , L. Campagna , and A. Siepel . 2022 . A deep-learning approach for inference of selective sweeps from the ancestral recombination graph . Molecular Biology and Evolution 39 : msab332 . OpenUrl CrossRef PubMed ↵ Hibbins , M. S. and M. W. Hahn . 2022 . Phylogenomic approaches to detecting and characterizing introgression . Genetics 220 : iyab173 . OpenUrl CrossRef PubMed ↵ Jackson , N. D. , B. C. Carstens , A. E. Morales , and B. C. O’Meara . 2017 . Species delimitation with gene flow . Systematic Biology 66 : 799 – 812 . OpenUrl CrossRef PubMed ↵ Kern , A. D. and M. W. Hahn . 2018 . The neutral theory in light of natural selection . Molecular Biology and Evolution 35 : 1366 – 1371 . OpenUrl CrossRef PubMed ↵ Kumar , V. , F. Lammers , T. Bidon , M. Pfenninger , L. Kolter , M. A. Nilsson , and A. Janke . 2017 . The evolutionary history of bears is characterized by gene flow across species . Scientific Reports 7 : 46487 . OpenUrl CrossRef PubMed ↵ Kutschera , V. E. , T. Bidon , F. Hailer , J. L. Rodi , S. R. Fain , and A. Janke . 2014 . Bears in a forest of gene trees: phylogenetic inference is complicated by incomplete lineage sorting and gene flow . Molecular Biology and Evolution 31 : 2004 – 2017 . OpenUrl CrossRef PubMed Web of Science ↵ Lan , T. , K. Leppälä , C. Tomlin , S. L. Talbot , G. K. Sage , S. D. Farley , R. T. Shideler , L. Bachmann , Ø. Wiig , V. A. Albert , J. Salojärvi , T. Mailund , D. I. Drautz-Moses , S. C. Schuster , L. Herrera-Estrella , and C. Lindqvist . 2022 . Insights into bear evolution from a Pleistocene polar bear genome . Proceedings of the National Academy of Sciences 119 : e2200016119 . OpenUrl CrossRef PubMed ↵ Li , H. 2013 . Aligning sequence reads, clone sequences and assembly contigs with bwa-mem . arXiv 1303.3997. ↵ Liu , S. , E. D. Lorenzen , M. Fumagalli , B. Li , K. Harris , Z. Xiong , L. Zhou , T. S. Korneliussen , M. Somel , C. Babbitt , G. Wray , J. Li , W. He , Z. Wang , W. Fu , X. Xiang , C. C. Morgan , A. Doherty , M. J. O’Connell , J. O. McInerney , E. W. Born , L. Dalén , R. Dietz , L. Orlando , C. Sonne , G. Zhang , R. Nielsen , E. Willerslev , and J. Wang . 2014 . Population genomics reveal recent speciation and rapid evolutionary adaptation in polar bears . Cell 157 : 785 – 794 . OpenUrl CrossRef PubMed Web of Science ↵ Long , M. , Z. Cao , J. Wang , and M. I. Jordan . 2018 . Conditional adversarial domain adaptation . arXiv 1705.10667. ↵ Mallet , J. , N. Besansky , and M. W. Hahn . 2016 . How reticulated are species? BioEssays 38 : 140 – 149 . OpenUrl CrossRef PubMed ↵ Miller , W. , S. C. Schuster , A. J. Welch , A. Ratan , O. C. Bedoya-Reina , F. Zhao , H. L. Kim , R. C. Burhans , D. I. Drautz , N. E. Wittekindt , L. P. Tomsho , E. Ibarra-Laclette , L. Herrera-Estrella , E. Peacock , S. Farley , G. K. Sage , K. Rode , M. Obbard , R. Montiel , L. Bachmann , Ó. Ingólfsson , J. Aars , T. Mailund , Ø. Wiig , S. L. Talbot , and C. Lindqvist . 2012 . Polar and brown bear genomes reveal ancient admixture and demographic footprints of past climate change . Proceedings of the National Academy of Sciences 109 : E2382 – E2390 . OpenUrl Abstract / FREE Full Text ↵ Mo , Z. and A. Siepel . 2023 . Domain-adaptive neural networks improve supervised machine learning based on simulated popu-lation genetic data . PLOS Genetics 19 : e1011032 . OpenUrl CrossRef PubMed ↵ Nielsen , R. , I. Hellmann , M. Hubisz , C. Bustamante , and A. G. Clark . 2007 . Recent and ongoing selection in the human genome . Nature Reviews Genetics 8 : 857 – 868 . OpenUrl CrossRef PubMed ↵ Pearce , T. , A. Brintrup , and J. Zhu . 2021 . Understanding softmax confidence and uncertainty . arXiv 2106.04972. ↵ Ray , D. D. , L. Flagel , and D. R. Schrider . 2024 . Introunet: identifying introgressed alleles via semantic segmentation . PLOS Genetics 20 : e1010657 . OpenUrl CrossRef PubMed ↵ Roux , C. , C. Fraisse , J. Romiguier , Y. Anciaux , N. Galtier , and N. Bierne . 2016 . Shedding light on the grey zone of speciation along a continuum of genomic divergence . PLOS Biology 14 : e2000234 . OpenUrl CrossRef PubMed ↵ Sanchez , T. , J. Cury , G. Charpiat , and F. Jay . 2021 . Deep learning for population size history inference: design, comparison and combination with approximate Bayesian computation . Molecular Ecology Resources 21 : 2645 – 2660 . OpenUrl CrossRef PubMed ↵ Schrider , D. R. , J. Ayroles , D. R. Matute , and A. D. Kern . 2018a . Supervised machine learning reveals introgressed loci in the genomes of Drosophila simulans and D. sechellia . PLOS Genetics 14 : e1007341 . OpenUrl CrossRef PubMed ↵ Schrider , D. R. , J. Ayroles , D. R. Matute , and A. D. Kern . 2018b . Supervised machine learning reveals introgressed loci in the genomes of drosophila simulans and D. sechellia . PLOS Genetics 14 : e1007341 . OpenUrl CrossRef PubMed ↵ Schrider , D. R. and A. D. Kern . 2018 . Supervised machine learning for population genetics: a new paradigm . Trends in Genetics 34 : 301 – 312 . OpenUrl CrossRef PubMed ↵ Sheehan , S. and Y. S. Song . 2016 . Deep learning for population genetic inference . PLOS Computational Biology 12 : e1004845 . OpenUrl CrossRef ↵ Smith , M. L. and B. C. Carstens . 2020 . Process-based species delimitation leads to identification of more biologically relevant species . Evolution 74 : 216 – 229 . OpenUrl CrossRef PubMed ↵ Smith , M. L. and M. W. Hahn . 2024 . Selection leads to false inferences of introgression using popular methods . Genetics 227 : iyae089 . OpenUrl CrossRef PubMed ↵ Solís-Lemus , C. and C. Ané . 2016 . Inferring phylogenetic networks with maximum pseudolikelihood under incomplete lineage sorting . PLOS Genetics 12 : e1005896 . OpenUrl CrossRef PubMed ↵ Strasburg , J. L. and L. H. Rieseberg . 2010 . How robust are “isolation with migration” analyses to violations of the im model? a simulation study . Molecular Biology and Evolution 27 : 297 – 310 . OpenUrl CrossRef PubMed Web of Science ↵ Taylor , G. A. , H. Kirk , L. Coombe , S. D. Jackman , J. Chu , K. Tse , D. Cheng , E. Chuah , P. Pandoh , R. Carlsen , Y. Zhao , A. J. Mungall , R. Moore , I. Birol , M. Franke , M. A. Marra , C. Dutton , and S. J. M. Jones . 2018 . The genome of the North American brown bear or grizzly: Ursus arctos ssp . horribilis. Genes 9 : 598 . OpenUrl CrossRef PubMed ↵ Tricou , T. , E. Tannier , and D. M. de Vienne . 2022 . Ghost lineages highly influence the interpretation of introgression tests . Systematic Biology 71 : 1147 – 1158 . OpenUrl CrossRef PubMed ↵ Wang , M.-S. , G. G. R. Murray , D. Mann , P. Groves , A. O. Vershinina , M. A. Supple , J. D. Kapp , R. Corbett-Detig , S. E. Crump , I. Stirling , K. L. Laidre , M. Kunz , L. Dalén , R. E. Green , and B. Shapiro . 2022 . A polar bear paleogenome reveals extensive ancient gene flow from polar bears into brown bears . Nature Ecology & Evolution 6 : 936 – 944 . OpenUrl CrossRef PubMed ↵ Wilson , G. and D. J. Cook . 2020 . A survey of unsupervised deep domain adaptation . ACM Transactions on Intelligent Systems and Technology (TIST) 11 : 1 – 46 . OpenUrl CrossRef ↵ Zou , T. , W. Kuang , T. Yin , L. Frantz , C. Zhang , J. Liu , H. Wu , and L. Yu . 2022 . Uncovering the enigmatic evolution of bears in greater depth: the hybrid origin of the Asiatic black bear . Proceedings of the National Academy of Sciences 119 : e2120307119 . OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted January 17, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following The reasonable effectiveness of domain adaptation for inference of introgression Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share The reasonable effectiveness of domain adaptation for inference of introgression Kerry A. Cobb , Megan L. Smith bioRxiv 2025.01.17.633659; doi: https://doi.org/10.1101/2025.01.17.633659 Share This Article: Copy Citation Tools The reasonable effectiveness of domain adaptation for inference of introgression Kerry A. Cobb , Megan L. Smith bioRxiv 2025.01.17.633659; doi: https://doi.org/10.1101/2025.01.17.633659 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Evolutionary Biology Subject Areas All Articles Animal Behavior and Cognition (7642) Biochemistry (17715) Bioengineering (13907) Bioinformatics (42003) Biophysics (21470) Cancer Biology (18624) Cell Biology (25533) Clinical Trials (138) Developmental Biology (13390) Ecology (19935) Epidemiology (2067) Evolutionary Biology (24356) Genetics (15617) Genomics (22529) Immunology (17753) Microbiology (40432) Molecular Biology (17200) Neuroscience (88681) Paleontology (667) Pathology (2840) Pharmacology and Toxicology (4828) Physiology (7653) Plant Biology (15171) Scientific Communication and Education (2046) Synthetic Biology (4304) Systems Biology (9826) Zoology (2271)