Genomic evidence for behavioral adaptation of herding dogs

Abstract

Herding dogs exhibit a distinct constellation of behaviors marked by inherent instinct and motor skills that manipulate and guide livestock in response to instructive commands and cues. Comparison of the whole-genome sequences of herding and nonherding breeds reveals signatures of positive selection associated with pathways underlying social interaction and cognitive functions. Of the strong selective sweep signals, haplotypes within ephrin type-B receptor 1 (EPHB1), which is linked to locomotor hyperactivity and spatial memory, show evidence of segregation within breed lineages for the conformation versus working lines of border collies and introgression with a genetically and geographically distant herding breed of Entlebucher mountain dogs. We show that a working line–specific haplotype of EPHB1 is associated with elevated levels of chase-bite motor patterns based on a well-validated behavior survey. These findings indicate that functional selection has shaped the genetic architecture of herding breeds, which may relate to their proficiency in addressing diverse tasks and challenges in maintaining control over the herd.

The whole-genome sequences of herding dogs reveal how selection has sculpted the unique behavioral traits.

INTRODUCTION

Domestic dogs (Canis lupus familiaris) display extraordinary variability in behavioral phenotypes, representing the selection and modification of traits needed for human survival (1). The primary development and diversification of these traits can be traced well before the 1800s when dogs were selected to perform functional tasks such as herding, guarding, retrieving, and hunting (2–4). Herding dogs share an innate physical and mental aptitude to move livestock. Generations of intensive selective breeding have amplified predatory instincts such as eye staring and chasing while effectively minimizing the natural inclination to kill prey (5, 6). The American Kennel Club (AKC) has assigned 32 breeds to the herding group, encompassing those with a strong impulse to work and excel in handling other animals (7). This group can be further divided into sheepdogs and cattle dogs, as defined by the Fédération Cynologique Internationale (fci.be/nomenclature/). Behavioral heterogeneity in herding dog breeds can be observed in their distinct herding patterns—circling and bringing livestock to the handler, driving livestock away in a trailing pattern, or stopping them by bringing them to the ground (8)—as well as individual behaviors, such as the intense stare known as “eye” together with “crouch.”

Although the availability of genomic tools has advanced genetic studies of a myriad of traits, the identification of genes underlying behavior patterns is poised to reveal the greatest number of surprises. The recent shared history of dog breeds with strong phenotypic and genotypic homogeneity offers unparalleled opportunities for biological discovery through interbreed mapping studies of morphological traits (e.g., skull shape, body size, and coat pattern) (9–13) and disease susceptibility. However, similar approaches have been ineffective in understanding the genetic foundation of an array of behavioral adaptations, including herding behavior, as the functions required of various breeds with common occupations are diversified and specialized (14). Behavioral adaptations provide additional complexity, as they are difficult to assess objectively with a high level of consistency and are heavily shaped by training, environment, and interactions with other living creatures, including humans (15). Most studies to date have focused on studies of anomalous behaviors, with a focus on those that are important to canine and human health, such as unprovoked aggression (16) or repetitive behaviors (17).

To address the longstanding challenges in identifying genetic factors contributing to herding behaviors, we harnessed the wealth of canine whole-genome sequencing (WGS) data from 12 herding breeds of different origins. These included dogs from Belgium: Belgian sheepdog, Malinois, Tervuren, and Bouvier des Flandres; the United Kingdom: border collie, bearded collie, Pembroke Welsh corgi, and Shetland sheepdog; France: Berger Picard; Australia: Australian cattle dog; Germany: German shepherd; and the United States: Australian shepherd (table S1). Each breed was treated independently to identify both the underlying breed-associated genetics and the common evolutionary paths associated with herding behaviors. To capture the genomic signatures of behavioral adaptation and historical forces that have shaped herding dog populations, we investigated targets of selection on the basis of herding breeds versus nonherding dog populations.

Selection for herding behavior has ancient roots, dating back to the early domestication of dogs, with historical records suggesting that the deliberate breeding of dogs by humans did not start until about 2000 years ago (18, 19), Thus, by integrating multiple methods, the cross-population composite likelihood ratio (XP-CLR) test (20) and the cross-population extended haplotype homozygosity (XP-EHH) test (21), we aimed to compile a more comprehensive catalog of likely selective sweeps and gain a deeper understanding of how selection has influenced dog variation. Specifically, the temporal sensitivity of these approaches varies: The XP-CLR test, which detects selection through changes in the frequency distribution between populations, is more effective at identifying older selection signals, whereas the XP-EHH test, which relies on extended linkage disequilibrium segments, is better suited for detecting signatures of recent selection (20, 22).

We further attempted to prioritize and substantiate the resulting sweep signals by (i) searching for allelic segregation within breed lineages within conformation versus working lines using standardized measurements from herding competitions, thus providing information about within-breed segregation, a product of functional selection; (ii) characterizing departures from expected haplotype sharing across genetically distant herding breeds; and (iii) using a well-validated behavioral survey to infer the phenotypic effects of candidate genetic variants. Overall, this study enhances our understanding of the evolutionary forces shaping the genomic makeup of behavioral adaptations found in the domestic dog genome.

RESULTS

Strong and shared selective sweeps across herding breeds

Each herding breed has been subjected to distinctive selective pressure spanning several generations, leading to the acquisition of desired behavioral characteristics that are either absent or only minimally present in nonherding dogs. The domestication-associated pressures resulting in canine herding behaviors offer a chance to investigate genomic alterations associated with behavioral traits. To identify the selective sweeps that may have arisen during phenotype-driven selective breeding, we took advantage of available WGS data from 551 dogs and 33 wolves (584 in total) with ~25× coverage of depth, yielding approximately ~50.2 million single-nucleotide polymorphisms (SNPs) and small insertions or deletions (indels) (table S1) (23).

In this study, the herding group comprises 12 AKC-recognized herding breeds, including the Australian cattle dog (n = 5), Australian shepherd (n = 5), Belgian sheepdog (n = 7), bearded collie (n = 7), Berger Picard (n = 6), Belgian Malinois (n = 5), border collie (n = 15), Bouvier des Flandres (n = 5), German shepherd dog (n = 15), Pembroke Welsh corgi (n = 5), Shetland sheepdog (n = 5), and Belgian Tervuren (n = 11) (table S1). For the nonherding group, we included only breeds that met two strict criteria: those classified into other breed groups (sporting, hound, working, terrier, toy, and nonsporting) and those ineligible for herding competitions. This group encompasses 91 breeds, with one representative dog randomly selected from each breed to constitute the control group (table S1). We aimed to identify genomic regions with extreme allele frequencies and linkage disequilibrium differentiation over extended linked regions using pairwise cross-population methods (XP-CLR and XP-EHH) (20, 21). The genomes of each herding breed were compared to nonherding dogs to characterize selective sweep signals. Using the empirical top 0.5% of genomic regions, a total of 2624 positively selected genes were detected in at least one herding breed (table S2).

For comparison, the same selective sweep analysis was independently performed on each of the four representative nonherding breeds of distinct origin, with a minimum sample size of eight, including the greyhound (n = 9), Portuguese water dog (n = 11), Tibetan mastiff (n = 10), and Yorkshire terrier (n = 74), using the same set of controls as in the herding breed analysis. Of the 2624 genes under selection in at least one herding breed, 2350 were retained after excluding overlapping sweep signals in nonherding breeds. Among the 2350 genes identified across multiple herding breeds, 471 genes (20.0%) were detected in at least two, 97 (4.13%) in at least three, and 23 (0.98%) in at least four herding breeds. These overlaps are significantly higher than what would be expected by random chance (permutation test, P < 2.2 × 10⁻¹⁶). To uncover potential clues about the common evolutionary forces shaping genetic variation patterns among herding dogs, we focused on the positively selected genes found in at least two breeds. These shared signatures of selection were considered candidates that have potentially undergone iterative modifications to optimize traits for canine herding.

Using the candidate gene set of shared signatures of selection, we tested for an enrichment of the number of genes that overlap loci previously linked to traits in the human National Human Genome Research Institute–European Bioinformatics Institute (NHGRI-EBI) genome-wide association study (GWAS) catalog (www.ebi.ac.uk/gwas/) (24). On the basis of the trait terms defined by this database, our analysis revealed that shared signatures of selection were significantly enriched for behavioral traits such as “cognitive function measurement,” “late-onset Alzheimer’s disease,” “unipolar depression,” “educational attainment,” and “intelligence” (permutation P < 0.05) (Fig. 1 and table S3). Of these, educational attainment and intelligence exceeded the multiple testing corrections [permutation false discovery rate (FDR) < 0.1]. To address the potential confounding effects of breed relationships, we also identified 367 genes under selection from at least two breeds originating in different countries, which are also significantly higher than expected by chance (permutation P = 8.32 × 10⁻⁸). The enrichment analysis of this robust gene set of shared signatures confirmed our initial findings, with educational attainment being the most significantly overrepresented trait (P < 0.005) (table S3). Herding dogs are agile and have been selectively bred to collaborate closely with humans to orchestrate livestock motion while avoiding harm to the livestock. Previously published behavioral tests on herding dogs demonstrate their exceptional inhibitory control (25) and human-directed behavior during problem-solving (25, 26), with the enriched traits demonstrating biological functions consistent with the expected occupational employment. For comparison, the positively selected genes in nonherding dog breeds showed overrepresentation for traits in different categories, such as the pulmonary function traits of the Portuguese water dog and the anthropometric traits of the Tibetan mastiff (Fig. 1 and table S3). This likely reflects the contrasting historical roles and distinct physical characteristics expected for nonherding dog breeds.

Fig. 1. Characterization of shared selective sweeps across herding breeds.

The significantly overrepresented traits by shared signatures of selection in herding dogs and other breeds based on the human GWAS catalog. The degree of enrichment (z-score) for each trait is color coded, and only the traits exceeding P < 0.05 are displayed. The traits satisfying FDR < 0.1 were indicated with a star. The traits are classified according to the experimental factor ontology (EFO) term definition. The enrichment analysis results for positively selected genes for each herding breed can be found in table S3. meas., measurement; BMI, body mass index; IGF-1, insulin-like growth factor 1; HbA1c, glycated hemoglobin; FEV/FEC ratio, forced expiratory volume/forced vital capacity.

Targets of selection in border collies

Border collies are a herding breed originating from counties along the border of Scotland and England, perhaps as early as the 1700s, and they have been referred to as “the most intelligent dog breed” for their ability to respond to a diverse range of commands and cues from human handlers (27). Border collies have been selected to specialize in the predatory sequence of orient, eye-stalk, chase, and grab-bite motor patterns during herding (28) and have consistently dominated competitive herding trials, functioning as the primary type of sheep herding dog in the world (29). Thus, we aimed to isolate the target of selection acting specifically on border collies by applying stringent thresholds to the sweep analysis (P < 0.001). We then retained candidate regions that contained border collie breed-specific SNPs in this dataset (Fisher’s exact test, P < 5 × 10⁻⁸) and may thus include causative variants in regions of positive selection. A comparison between border collies and nonherding dogs revealed 24 genes that exceeded the stringent significance threshold (Fig. 2, fig. S1, and table S4). Among these, ephrin type-B receptor 1 (EPHB1) was among the strongest, showing significant signatures of selection (both XP-EHH and XP-CLR in the top 0.1%) and containing the border collie–specific T allele (76%) at an intron variant (chr23:31592984), which was absent in 422 nonherding dogs from 91 nonherding breeds (0%) (Fig. 2, A and B). The strongest selective sweep included at least eight genes previously suggested as regulators of cognitive function, including memory retention (CD44), learning (COL25A1), spatial memory (RHOT1, EPHA5, and EPHB1), and motor learning (PEX5L), and social behaviors including aggression (EPHA5), social interaction (IRX2), coping behavior (PEX5L), addictive behavior (TFAP2B), and mood disorder (CD44) (table S4). All of these can now be considered candidates for functional study in dogs. Using publicly available RNA sequencing (RNA-seq) data from 245 dog samples comprising 15 tissues, we found that 11 of 24 positively selected genes (Fig. 2C and fig. S1) were highly expressed in brain tissues [average transcripts per million (TPM) > 10], which were more prominent when compared to other tissues such as the skin, liver, kidney, pancreas, adrenal gland, and testes, indicating their functional roles in nervous system processes.

Fig. 2. Identification of EPHB1 as a target of selection in border collies.

(A) The selective sweep signals (XP-EHH and XP-CLR) around the EPHB1 peak with the genome-wide significance level. (B) The barplot shows significant allele frequency differentiation between border collies and all the nonherding dogs within the EPHB1 gene. (C) Boxplot of TPM for EPHB1 gene across tissues from publicly available RNA-seq data.

Allelic segregation between working and show lines of border collies within the EPHB1 gene

For many dog breeds, including border collies, there exist both “working” and “conformation” lines. The former are bred for occupation and are kept in the breeding pool if they excel at their given tasks. The latter are bred for adherence to the stated breed standard and exhibit specific morphologic features, e.g., coat color, facial feature, ear position, etc. (30, 31). We hypothesized that the differential selective pressure acting on lineages associated with the two types of border collie might be genetically discernible. The border collie is a particularly interesting breed in this context, as it was formally recognized by the AKC in 1995 and thus became eligible to participate in conformation competitions only very recently. This recent split into working and conformation lines could argue that strong selective pressures on small numbers of genes were sufficient to develop these lineages. To identify the underlying loci, genotypes of the border collie working (BORDW; n = 50) and border collie conformation (BORDC; n = 9) dogs were generated using the Affymetrix Axiom Canine HD Array (670,000 SNPs). The BORDW category pertains to individuals formally registered with the American Border Collie Association, an organization specializing in validating pedigrees for border collies bred as working dogs. Conversely, BORDC exclusively encompasses dogs that have attained conformational titles through participation in national-level AKC-sponsored competitions, denoting a focus on their ability to present specific physical attributes. We first explored these two populations using a Bayesian approach and a set of autosomal SNPs (32). Assuming two source populations (K = 2), the results reflect the genetic separation of the BORDW and BORDC lineages (Fig. 3A).

Fig. 3. Segregation between working and show lines of border collies.

(A) The proportion of ancestry for each BORDW and BORDC individual assuming two ancestral populations (K = 2). (B) A linear mixed model was applied to assess genetic associations, identifying genes that significantly distinguish the populations. Blue points represent SNPs that satisfied the Bonferroni correction (P = 1.10 × 10⁻⁷), while red points indicate the top five most significant genes. There was no evidence of genomic inflation due to population substructure (λ = 0.995).

We then conducted a GWAS to identify causative variants differentiating the 50 BORDW and 9 BORDC described above using 452,926 SNPs after applying a minor allele frequency threshold of 0.01. To correct for population structure, GWAS was performed in genome-wide efficient mixed model association (GEMMA) (33) using a kinship matrix and linear mixed model. There was no evidence of genomic inflation due to population substructure (λ = 0.995) (Fig. 3B). A total of 176 genome-wide significant SNPs exceeding Bonferroni significance (P = 1.10 × 10⁻⁷) were identified. The top five most significant loci include the protein-coding genes PTPRZ1, ASB3, LMO3, LRP2, and EPHB1 (Fig. 3B), the last of which was also identified in the selective sweep analysis. Although none of these have been previously associated with canine traits (table S5), the relevant pathways have been. For instance, a previous report associated multiple ephrin receptor genes, including EPHA5 and EPHA3, important in axon guidance, with border collie lineage (34). The identification of the EPHB1 gene within the common ephrin signaling pathway underscores the pronounced prevalence of loci associated with synaptic components, indicating their potential relevance to a multitude of canine behavioral processes.

Moreover, among the genome-wide significant genes were MSX1 and RUNX2, which are notable for their roles in craniofacial development and morphogenesis in mice (35, 36), suggesting that the selective pressures on working lines to create conformation lines included pressure to genetically fix aspects of morphological traits such as skull shape. We speculate that at least some of the alleles observed could reflect selection for behavioral modifications that made the BORDC dogs more suited to compete in the show ring, displaying less herding-like tendencies and a quieter and more cooperative stance. Overall, these genes highlight the critical pathways that have been under selection to create distinct lineages of border collies in a very short period of time and should be functionally explored for their potential role in affecting behavior and morphology.

Significant haplotype sharing across herding breeds within the EPHB1 gene

The presence of distinct haplotype combinations across phylogenies indicates hybridization and introgression (37). To track the genetic signatures and enhance the analysis with haplotype sharing for a robust assessment of canine population structure within this region, we constructed haplotypes in eight populations using the available whole-genome sequences of the Entlebucher mountain dog (EMD; n = 8), Yorkshire terrier (n = 74), rottweiler (n = 10), Tibetan mastiff (n = 10), Portuguese water dog (n = 11), and gray wolf (n = 33); and SNP chip data of BORDW and BORDC. The common variants between WGS and SNP array data were used to infer haplotypes. The analysis revealed a total of six distinct haplotypes spanning ~57 kb in the EPHB1 gene from the eight tested populations. We found that the frequency of haplotype no. 1 was close to fixation in BORDW (84%), notably less frequent in BORDC (22%), and nearly absent in other nonherding and herding breeds. Still, the EMD had a frequency of 63% (Fig. 4A). At this haplotype, BORDW and EMD clustered together and were distinct from the BORDC and the remaining two Swiss mountain dog populations (Greater Swiss and Bernese mountain dogs with n = 2 and 18, respectively) and wolves (Fig. 4A). BORD (UK sheepdog) and EMD (Swiss cattle dog) are geographically and genetically distant at the genome-wide level (Fig. 4B). To further test whether haplotype sharing within EPHB1 across different clades could be expected by chance, we estimated the four-taxon D statistics (fd) (38) in 200-kb sliding windows across the genome, incorporating BORDC and wolf populations, which estimates an excess of shared derived variants between BORDW and EMD. We found a significantly higher gene flow between BORDW and EMD within this region, a characteristic footprint of local introgression, as the fd value of this region was in the top 0.5% of the empirical distribution (P < 0.005) (Fig. 4C).

Fig. 4. EPHB1 haplotype construction and evidence of introgression.

(A) Haplotypes within the EPHB1 gene region surpassing frequency of 20% in at least one breed are defined and indicated. All rare haplotypes (<20%) are merged. (B) F_ST (fixation index)-based phylogeny among BORDW, BORDC, wolf, and three Swiss mountain dogs (Entlebucher, Greater Swiss, and Bernese) according to the genome (left) and the window encompassing the region defined in [(A), right]. (C) The same region within EPHB1 was tested for significant introgression between BORDW and Entlebucher mountain dogs in contrast to all 50-kb windows at the genome-wide level (empirical P < 0.005). 3′UTR, 3′ untranslated region; 5′UTR, 5′ untranslated region; YORK, Yorkshire terrier; ROTT, rottweiler; TIBT, Tibetan mastiff; PWD, Portuguese water dog; GSMD, Greater Swiss mountain dog; BMD, Bernese mountain dog. fdM, modified four-taxon D statistic.

EPHB1 controls the preoptic regulatory factor-2 and is involved in axon growth and spatial memory (39). Genetic deletion of EPHB1 in mice leads to neuronal loss within the substantia nigra and spontaneous locomotor hyperactivity (40). EPHB1 knockout mice show aberrant thalamic-cortical axon guidance (41), and the gene is one of many included in deletions associated with Williams-Beuren syndrome, a rare human neurodevelopmental disorder characterized by hypersociability (42). The presence of the EPHB1 receptor is also involved in the perception of neuropathic pain and the establishment of physical dependence on morphine, indicating its potential role in these important processes (43). Together, these results suggest that during the development of herding breeds, selection targeted unique genomic regions such as EPHB1, and the sharing of this adaptive phenotype presents a model for investigation of potentially shared beneficial alleles and their role in herding.

Genetic effect of EPHB1 on behavioral traits from owner-directed survey

Owner-centered survey methodology is a frequently used tool for evaluating canine behavioral traits (44). These surveys leverage the knowledge and insights of dog owners and caregivers when assessing the behaviors of their own dogs. Aggregating responses from numerous independent owners and caregivers make it possible to mitigate individual biases (45) and are highly valuable in the investigation of behavioral distinctions specific to different dog breeds (14). To provide a clearer picture regarding the effect of the working line–specific haplotype within EPHB1 on behavior, we examined its association with publicly available owner-reported behavioral phenotypes (46). The Darwin’s Ark survey data consisted of 117 questions, reduced the dimension to eight factors (human sociability, arousal level, toy-directed motor patterns, biddability, agonistic threshold, dog sociability, environmental engagement, and proximity seeking), and scored 2155 dogs that have also undergone low-pass genome sequencing (table S1). Of the eight factors tested, we found that the dogs carrying the working line–specific haplotype (n = 337) were significantly more likely to be toy directed than were the noncarriers (n = 1535) (P = 4.3 × 10⁻⁴) (Fig. 5 and table S6). The toy-directed motor pattern was characterized by a propensity to engage with toys and objects, manifesting as grab-bite and chase behaviors. We found congruent results when we only investigated mixed breeds from Darwin’s Ark (n = 179 dogs versus 893 dogs) (P = 4.1 × 10⁻⁴) and even within border collies (n = 41 dogs versus 58 dogs) (P = 1.9 × 10⁻²), indicating that the association is robust against owner bias of breed stereotypes (fig. S2). The two analyses incorporating all individuals and mixed breeds, respectively, showed strong evidence of association even after Bonferroni multiple testing corrections for 24 multiple tests (8 behavior factors × 3 analyses, threshold P = 2.1 × 10⁻³). The association within the border collie population only satisfied the nominal threshold, likely due to low statistical power. Overall, we identified haplotypes within a gene critical for the neuronal function that distinguishes border collie lines selected for strong herding behaviors versus show-quality border collies, as well as other nonherding breeds. Further, the association of key haplotypes with dog behaviors such as toy-directed motor patterns and grab-bites and chases, while not unique to herding breeds, is significant and may provide a ready assay for a functional investigation into these results.

Fig. 5. Genetic effect of EPHB1 on behavioral traits from owner-reported survey.

The individuals with the border collie working line–specific haplotype within EPHB1 (haplotype no. 1) were significantly more likely to be toy directed than the noncarriers on the basis of the owner-directed survey data. The congruent results incorporating only mixed breeds (n = 179 dogs versus 893 dogs) (P = 4.1 × 10⁻⁴) and only border collies (n = 179 dogs versus 893 dogs) (P = 4.1 × 10⁻⁴) are shown in fig. S2.

DISCUSSION

A long-standing objective of canine geneticists has been to identify genes responsible for behavioral characteristics associated with specific breeds, such as those observed in herding breeds. The human-canine relationship has evolved to a sophisticated state, characterized by the selection of dogs that epitomize refined traits tailored for specialized aspects of herding. Each herding behavior entails the precise execution of motor actions, the individual’s innate drive to tackle challenging tasks, and their capacity to discern and decipher subtle alterations in human cues and livestock body language (28, 47).

One approach for investigating canine adaptive behavior is the comparison of dog breeds instead of individual dogs, relying on breed group–level behavioral characteristics. We have previously illustrated how strong behavioral selection can modify physiology to create breeds with superior speed and athleticism by a genome-wide comparison of the sport hunting to the terrier breeds, groups at the end of a continuum in both form and function (48). However, the substantial range of behavioral differences observed among herding breeds emphasizes the necessity for comprehensively considering both local and global complexities inherent in various relationships, thereby avoiding bias toward any specific category (49). These relationships encompassed distinctions between breeds (herding versus nonherding), within-breed segregation (conformation versus working), across genetically distant breeds sharing similar occupations (e.g., border collies and Entlebucher mountain dogs), and even within individuals (samples with and without putative working–associated haplotypes).

Our findings indicate that the EPHB1 gene met all the tested differentiations, potentially underscoring its role in driving the adaptive performance capabilities observed among different dog breeds during domestication. In a recent study, Dutrow et al. (34) demonstrated that behavior has played a role in the trajectory of canine genetic diversification by aligning canine lineages with historical breed occupations and using breed-average behavior metrics from Canine Behavioral Assessment and Research Questionnaire (C-BARQ). They found that genes linked to lineages converge in neurodevelopmental pathways, revealing an enrichment of axon guidance functions associated with sheepdogs. The correlated signatures observed in the interrelated axon guidance pathways with the current findings, specifically involving the ephrin receptor EPHA5 gene, highlight the importance of this molecular pathway, as does our identification of the EPHB1 gene. Together, these strong candidate genes suggest selection on a multigenic trait architecture and parallelism primarily through pathways rather than repeated single gene effects (50).

Limitations of the study

Because of dog population structure and breed history, the study has some natural limitations. Herding breeds exhibit distinct behavior patterns tailored to their roles in managing livestock, leading to specialized styles: e.g., herders (e.g., border collies), heelers (e.g., Australian cattle dog), boundary dogs (e.g., German shepherd dogs), and drovers (e.g., Bouvier des Flandres) (table S1) (28, 51). However, the genetic basis of adaptive behavior may show convergence among herding breeds with similar specialized tasks, suggesting interesting hypotheses about distinct genetic adaptations linked to specific herding styles and environments. These hypotheses can be rigorously tested but will require much larger sample sizes than those used here.

This study relies on breed assignment for assessing phenotype, at least with regard to the identification of selective sweep sharing among herding breeds. Thus, our results are focused on identifying pathways and a small set of genes rather than nucleotide-level variants and do not address the issue of structural variants. More refined studies would require individual-level behavior metrics on each dog and, again, larger numbers of samples. However, our identification of social and neural function pathways under selection sets the stage for such studies, by providing both a starting point for in-depth studies of herding behavior and by defining the types of behavioral assays needed.

Last, while genomic evidence for selection is suggestive, it is not conclusive. Robust evidence that a specific locus influences a trait requires multiple independent functional validation experiments, as identification of significant alleles does not provide evidence of function, regardless of how provocative a perceived causal relationship is (52, 53). Incorporating data on qualitative and quantitative variations in proteins and metabolites, along with epigenetic modifications, will ultimately provide a more comprehensive understanding of the impact of genetic perturbations on the organism as a whole (54).

The behavior patterns observed in modern dog breeds reflect historical needs for human survival, and herding dogs display a rich and varied system for genetic studies of such behaviors. In this study, we observe that signatures of selection in herding dogs are highly enriched in genes relating to social interaction and cognitive functions, which can be putatively linked to a heightened sensitivity to human communication, an ability to quickly grasp new commands, and, last, proficiency in addressing a wide array of tasks and challenges in maintaining control over the herd. In essence, herding dogs’ unwavering commitment to rounding up livestock and eagerness to please humankind highlights the genomic footprint of adaptations within domestic dogs.

MATERIALS AND METHODS

Sample data

The WGS data used in this study were collected from the National Center for Biotechnology Information Sequence Read Archive (SRA) (www.ncbi.nlm.nih.gov/sra). Samples identified by accession numbers PRJNA448733, PRJNA576632, PRJNA685036, and PRJNA726547 were used, comprising 551 canine and 33 wolf samples. Samples from domestic dogs were collected with the signed consent of their owners. The sample collection process involved no direct interaction with the dogs and was conducted with institutional approval under NHGRI Institutional Animal Care and Use Committee (IACUC) protocol GFS-05-1. To generate the variant call format (VCF) file, the produced sequence reads were mapped against the CanFam 3.1 reference genome (genome.ucsc.edu/cgi-bin/hgGateway?db=can-Fam3). Our variant calling methodology strictly followed the protocol established in prior studies, including the recently published Dog10k (23, 55), encompassing steps from sequence read trimming and alignment to variant identification and refinement (56–61). Through this process, approximately 50.2 million high-quality autosomal SNPs and indels were called and used for downstream analysis.

Selective sweep detection

We used XP-EHH to detect selective sweeps (https://github.com/szpiech/selscan) (21). The XP-EHH score is directional; the higher the value in the positive direction, the stronger the selective sweep in herding dogs. We defined each analysis window as 50 kb, ensuring no overlap between windows, and selected the highest XP-EHH score in each window as the representative value. To account for varying single-nucleotide variant (SNV) densities in each window, a binning strategy was used on the basis of SNV count: Windows were grouped into bin1 (1 to 100 SNVs), bin2 (101 to 200 SNVs), bin3 (201 to 300 SNVs), and bin4 (more than 300 SNVs). Subsequently, an empirical P value was obtained according to the XP-EHH score rank within each bin, and a window satisfying P < 0.005 (top 0.5%) was designated as the potential sweep region (62).

We also performed an XP-CLR test (https://reich.hms.harvard.edu/software) to detect selective sweeps that jointly model multilocus allele frequency between the two populations (20). For this analysis, we used parameters including discrete, nonoverlapping 50-kb windows, limiting to a maximum of 600 SNVs per window and applying a correlation threshold that reduced the influence of SNVs on the XP-CLR outcome to 0.95. The window corresponding to the top 0.5% of the calculated XP-CLR scores for each window was designated as the potential sweep region. For both selection scans, a more stringent threshold was applied as opposed to the more commonly used 1% in recent canine studies (23, 48, 63), to reduce the incidence of false positives arising from random genetic drift. Significant genomic regions identified from each statistic were annotated to the closest genes (±10 kb) for each herding breed. To identify border collie–specific SNPs, as found in this dataset, Fisher’s exact test was performed to compare allelic segregation against nonherding dog samples using PLINK software (version 1.9) (64).

GWAS enrichment analysis

SNPs associated with human traits were obtained from the NHGRI-EBI GWAS catalog (www.ebi.ac.uk/gwas/) (24). These SNPs were mapped onto the dog genome (CanFam3.1) using the UCSC lift genome annotation (https://genome.ucsc.edu/cgi-bin/hgLiftOver). Only traits with at least 300 associated variants were considered in subsequent enrichment analysis. For each trait, the total number of XP-EHH or XP-CLR significant regions that overlapped at least one GWAS variant was calculated. A permutation approach was then applied to assess whether this overlap was higher or lower than expected by chance. In each permutation, an equivalent number of genomic regions of the same sizes as the positively selected regions were randomly sampled across the dog genome to count the number of regions overlapping at least one GWAS variant. This process was repeated 1000 times to generate a null distribution of overlap counts for each trait. The mean number of overlaps from these randomized iterations was then used to compute a z-score and the corresponding P value, allowing us to assess whether the observed overlap was significantly higher than expected by chance. The same approach was applied to test the number of overlaps between positively selected genes in each breed, with traits classified according to the experimental factor ontology (EFO) term definition (table S3).

Genome-wide association analyses

DNA samples from 50 BORDW and 9 BORDC dogs were genotyped at 670,000 loci using the Affymetrix Axiom Canine HD Array. Sample information is provided in table S1. The SNP genotype dataset has been deposited in figshare and is available at https://figshare.com/s/413f62bd3f89d97d3fcc. Pruning for minor allele frequency < 0.01 in the population yielded 452,926 variants. Association analyses were performed using GEMMA (33). To infer the population structure within border collies, ADMIXTURE was used with the assumption of two ancestral populations (K = 2) (65). Haploview software (66) was used to evaluate the haplotype structure and estimate haplotype frequencies within the EPHB1 gene. The PLINK (“--recode HV” function) was used to convert from the PLINK file to create a Haploview format for use in the haplotype construction.

Owner-reported survey data

The Darwin’s Ark survey (46) included 117 questions distilled into eight factors: human sociability, arousal level, toy-directed motor patterns, biddability, agonistic threshold, dog sociability, environmental engagement, and proximity seeking. These factors were used to score 2155 dogs, all of which had also undergone low-pass genome sequencing (table S1). To infer the phenotypic effect of candidate genetic variants, we used the publicly available Darwin’s Ark genetic dataset of 2155 individuals in VCF (https://data.broadinstitute.org/DogData/). The haplotypes within the EPHB1 were determined using the Beagle version 4.1 (67). We used the t test statistic to examine the behavioral score difference between the dogs based on the presence and absence of candidate alleles/haplotypes.

Introgression between border collie and Entlebucher mountain dog

The ABBA-BABA test was used to detect admixture events by comparing the frequency of incongruent site patterns (ABBA and BABA) across four populations (BORDW, BORDC, EMD, and wolf samples) with statistical significance assessed using a block jackknife resampling procedure (38). Dsuite software was used with the following parameters: “Dsuite Dinvestigate -w 25,5” (68).

RNA-seq data analysis

RNA-seq data from 231 samples were obtained from the SRA (www.ncbi.nlm.nih.gov/sra) (table S1). The quality of the RNA-seq data was evaluated using FastQC (www.bioinformatics.babraham.ac.uk/projects/fastqc/; version 0.11.9), followed by the removal of low-quality reads using Trimmomatic (version 0.39) (69). Mapping to the CanFam 3.1 reference genome was performed using the STAR aligner (version 2.7.3a) (70). Expression levels were quantified to produce TPM values, using the rsem-calculate-expression feature of RSEM (71).

Supplementary Materials

The PDF file includes:

Figs. S1 and S2

Legends for tables S1 to S6

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S6