Genetic testing predicts appearance but not behavior in dogs

Significance

Using data from Darwin’s Ark, a community science project with over 3,000 dogs with genetic and behavioral data, we critically evaluate reported behavior-associated genetic variants. We find no evidence that these variants predict behavior. Some variants are correlated with aesthetic traits that define breeds, suggesting that the earlier breed studies that linked these variants to behavior were confounded by the complex dog population structure. Genetic tests focusing on a few variants are unlikely to provide accurate predictions for polygenic behavioral traits or complex diseases in dogs. Instead, large, carefully designed studies that consider the whole genome are essential for developing predictions that can improve dog health and welfare.

Abstract

Genetic tests for behavioral and personality traits in dogs are now being marketed to pet owners, but their predictive accuracy has not been validated. To evaluate the reliability of such tests, we analyzed data from Darwin’s Ark, a community science initiative that includes over 3,000 dogs with both genetic data and individual-level behavioral phenotypes. None of the candidate variants had significant associations or predictive power for behavioral traits as previously reported. However, we found strong associations with aesthetic traits that differentiate breeds, such as height, leg length, and ear shape. Our results suggest that earlier studies using breed-average phenotypes, rather than individually measured phenotypes, were confounded by population structure. Behavior in dogs is polygenic and complex, and thus cannot be accurately predicted using tests that consider only a few genetic variants. Furthermore, behavior in dogs is only moderately heritable, and environmental influences inherently limit the potential accuracy of genomic predictions. Developing meaningful, accurate genetic predictions for complex traits that can improve dog health and welfare will require very large cohorts of individually phenotyped dogs.

Genome-wide association studies (GWAS) are a powerful tool for identifying genetic variants, typically single-nucleotide polymorphisms (SNPs), associated with traits or diseases. GWAS are effective for both simple traits caused by mutations in a single gene and complex traits influenced by many genetic and environmental factors. In humans, GWAS have revealed novel pathways underlying complex diseases and opened new avenues for therapeutic development (1). The power of GWAS is largely determined by sample size, and the largest GWAS include millions of individuals (2).

To estimate an individual’s risk of exhibiting a complex trait, researchers use polygenic risk scores (PRS), which aggregate the effects of many genetic variants across the genome discovered through GWAS. The potential for accurate prediction depends on both the heritability of a trait, meaning the proportion of variation explained by genetic factors, and its polygenicity, or the number of genetic variants that influence it. Traits with lower heritability require much larger sample sizes, and traits with very low heritability cannot be predicted accurately from genetic information alone (3, 4).

GWAS has been widely used in dogs to identify genomic loci associated with aesthetic traits, like size and color, behavioral traits, and disease risk. The history of dog breeding makes GWAS in dogs unusually powerful. About 200 years ago, humans began purposeful breeding of lineages of dogs with unusual physical characteristics (5), creating hundreds of modern breeds (6–11). The limited genetic diversity within dog breeds enhances the power to detect associations in GWAS (12, 13). However, this history also exacerbates issues of population structure known to confound genetic studies by causing both false positives (14) and missed true associations (15).

To date, genetic prediction in dogs has been limited to simple traits and highly heritable complex traits. For example, genotyping a dog at a single locus can determine whether they have an inherited form of progressive blindness (2, 3). Aesthetic traits under selection in breeds are often exceptionally heritable and explained by a small number of large-effect loci, making them particularly amenable to genetic prediction. We developed a highly accurate PRS for height, which is more than 99% heritable, by combining information from 2,733 SNPs identified in a GWAS with 1,730 dogs (16, 17).

Currently, the largest GWAS in dogs have fewer than 10,000 individuals (16, 18), making developing PRS for less heritable traits infeasible (19). This includes behavioral and personality traits in dogs, which are only weakly or moderately heritable in studies that use paired phenotype and genetic data from individual dogs (11, 16, 20–22). Many factors other than genetics influence behavior, including training, socialization, housing, and early-life experiences. Developing accurate PRS in dogs for such traits will likely require studies with tens or hundreds of thousands of dogs (23–26).

To circumvent the sample size problem, some studies combine behavioral scores from one set of dogs with genetic data from a different set of dogs. To test for association, dogs with genetic data are assigned a phenotype based on the average score of dogs in their breed (27–34). This approach has been effective when applied to aesthetic traits that have very little variability within breeds (35). However, behavioral traits in dogs vary nearly as much within breeds as between breeds (16, 36–39). Furthermore, the true sample size of these studies is much smaller than conventional genetic association studies. While they include large numbers of dogs, their statistical power is ultimately constrained by the number of independent units of analysis (the number of breeds), and the analysis vulnerable to pseudoreplication if nonindependent observations are treated as if they were independent (40). In this manuscript, we use the term GWAS exclusively to describe studies that pair genetic data with individual phenotype data, consistent with common terminology, and “breed-average studies” to describe those using population-level phenotypes.

To date, breed-average genetic studies have reported hundreds of SNPs associated with behavioral traits (27–34) conflicting with results from GWAS, which have identified only a few significant associations (16, 21, 22, 40, 41). Despite this, genetic tests based on these studies are being marketed to consumers (42, 43). For example, if a dog is labeled as genetically predisposed to aggression, an owner might limit essential social interactions, or a shelter might decide against adoption.

Here, we use Darwin‘s Ark, a large community science project collecting individual phenotypes and genetic data (44), to investigate whether genetic variants from breed-average studies can accurately predict behavior in pet dogs from diverse breeds, mixed-breed dogs, and nonbreed dogs.

Results

Identifying 151 Candidate Behavior SNPs from Previous Studies.

We reviewed recent studies that relied on breed-average phenotypes to identify candidate SNPs associated with behavior in dogs (27–30, 32–34, 45). These studies have reported hundreds of significantly associated SNPs, and several have proposed that behavior and personality traits of both breeds (27) and individual dogs (28) can be predicted from these variants. However, inferring individual-level associations from group-averaged data is statistically problematic (46). In genetic studies, individual-level data are essential for controlling for relatedness, population stratification, and other factors that lead to biased and spurious results (47, 48). These issues are especially pronounced in dogs due to high stratification between breeds and close relatedness within breeds (49).

Four of the studies scored behavioral traits that are also measured in Darwin’s Ark (SI Appendix, Table S1) (27–29, 34, 50). In each of these studies, the average score on each trait for each breed was calculated using survey data from dogs without accompanying genetic data (51), and those average scores were used to assign phenotypes to unphenotyped dogs genotyped or sequenced for other studies. This included genome-wide studies of (1) 14 traits in 101 breeds (34); (2) canine fear and aggression traits in 29 breeds across two published datasets (27); (3) 17 behavioral traits using genetic data from three studies with 29 to 50 breeds (29). The fourth study used SNPs from the previous three studies to predict problem behaviors in 397 dogs (28). In total, we identified 151 candidate behavior SNPs (SI Appendix, Table S2).

Many of the SNPs proposed to predict behavior in dogs occur near genes known to influence aesthetic traits that differentiate dog breeds and are artificially selected for by breeders (SI Appendix, Table S3) (30, 52–83), suggesting that population structure rather than associations with behavioral traits may explain these associations (15, 84, 85). For example, 18 of the 20 SNPs proposed to predict behavior in individual dogs in a recent publication (28) overlap loci associated with breed-defining traits: nine loci associated with size, including near the major size-associated genes IGF1 (2 SNPs), IGF2BP2 (1 SNP), HMGA2 (2 SNPs), and MSRB3 (2 SNPs) (16, 30, 74, 75, 77); five loci associated with skeletal morphology (73, 74, 76); and four associated with fur texture and coat color, including at RSPO2 (2 SNPs), FGF5 (1 SNP), and MITF (1 SNP) (53, 59, 73, 74).

No Evidence of Association with Behavior in Pet Dogs.

With the advent of large-scale community science in pets, we now have the capacity to do GWAS in dogs as a more powerful and less problematic alternative to breed-average studies. GWAS require individual-level phenotype and genetic data but are an exceptionally powerful tool for discovering variants associated with complex traits (1). Through our online community science platform Darwin’s Ark (www.darwinsark.org) we have collected both phenotypes and genetic data for thousands of dogs (16). Darwin’s Ark is a nonprofit, open data initiative that engages dog and cat owners in research by inviting them to complete short surveys and offering an opportunity to sequence their pet’s genome. In previous work, we showed that owner survey responses align with those from dog experts and are not strongly influenced by cultural breed stereotypes (SI Appendix, Supporting Text) (16).

In contrast to studies that rely on breed-average phenotypes, Darwin’s Ark includes both single-breed and mixed-breed dogs in proportions similar to the US pet dog population (16), making it ideal for assessing predictive tests marketed to dog owners. We generated high-density, highly accurate genotype data for 3,287 dogs enrolled in Darwin’s Ark through low-coverage sequencing (mean coverage = 0.82x) and imputation using reference genomes from Dog10K (86) (SI Appendix, Supporting Text). In total, we call genotypes for 29,089,701 autosomal SNPs. We estimated each dog’s breed ancestry using global admixture analysis, comparing their genotype to a reference panel of 1,212 dogs from 101 different breeds, following previously published methods (16). Only a minority of dogs (25%) have ancestry from just one breed (single-breed or “purebred” dogs) (SI Appendix, Table S4), and most (53%) have discernible ancestry from at least four breeds (Fig. 1A). The dogs span a wide range of ages, from infant to 22 years of age, with a median age of five years (IQR=6.7) (Fig. 1B). The most common breed ancestry is American pit bull terrier (11.7%) followed by Labrador retriever (6.8%) and Chihuahua (6.4%) (Fig. 1C).

Fig. 1.

No evidence for association with candidate behavior SNPs in a population of 3,295 pet dogs. (A) Dogs vary widely in size, and most have ancestry from more than one breed. About 1 in 4 (24.4%) are single-breed (“purebred”) dogs. (B) Ages range from 0 to 22 y (mean 5.7 ± 4.0 SD), with an even sex distribution (49.9% female; 93.8% spayed or neutered). (C) The 15 most common breed ancestries, drawn from at least 85 breeds contributing at least 5% ancestry to at least one dog. (D) In GWAS, some of the candidate behavior SNPs are significantly associated with aesthetic traits (green), but none are associated with any of the seven behavior traits (orange). The vertical dashed line marks the genome-wide significance threshold (P < 5 × 10⁻⁸); significant associations are labeled by trait. (E) Aesthetic traits are more heritable than behavior traits. SNP-based heritability estimates (h²) are higher for aesthetic traits (green) than for behavioral traits (orange). Points show results from LD-corrected (open circles, Top) and uncorrected GREML (solid circles, Bottom). Boxplot shows the median (horizontal line), interquartile range (box), and whiskers extending to the smallest and largest values within 1.5 × IQR of the lower and upper quartiles. (F) All behavioral traits (orange points) are more variable within breed than between breeds, but some aesthetic traits (green points with labels) are less variable within breeds than between breeds. The fraction of variability between breeds is strongly correlated with trait heritability (Pearson test; r = 0.92, P = 5 × 10⁻¹⁹).

The Darwin’s Ark survey data include 280 questions, from which we previously derived eight behavioral factors (16). Of these, 34 questions and four behavioral factors correspond to descriptions of seven of the fourteen C-BARQ traits used in the four breed-average studies (SI Appendix, Tables S1 and S5). Another ten questions measure aesthetic characteristics that differentiate breeds.

We ran GWAS in the current Darwin’s Ark dataset for all questions that overlapped with the breed-average studies, all eight behavioral factors and all aesthetic traits. This increased our sample size by about 1.4× (1,794 to 3,074 dogs per trait) compared to our previous dataset, including six additional behavioral traits and four additional aesthetic traits (SI Appendix, Table S5). We controlled for population structure and cryptic relatedness using a mixed linear model approach incorporating a genetic kinship matrix. We included age, sex, and spay/neuter status as covariates in all GWAS, and size as a covariate in all but the size GWAS. In total, we identified 6596 SNPs significantly associated with 20 aesthetic traits (6570 SNPs) and 7 behavioral traits (26 SNPs).

The inclusion of mixed-breed dogs in our GWAS increases our statistical power. We compared GWAS in a cohort of dogs with exclusively mixed-breed ancestry (1,764 dogs with no more than 45% ancestry from any single breed) to GWAS in the same number of dogs randomly sampled from the full cohort, in which about 40% of the dogs are either single-breed or have at least half their ancestry from one breed. A majority of SNPs (61%; 690/1,132) that were genome-wide significant in the full cohort showed stronger associations in the mixed-breed subset than in the random subset (Wilcoxon P = 1.6 × 10^–19). This is consistent with lower pairwise relatedness in the mixed-breed cohort increasing effective sample size, and thereby improving GWAS power.

Our GWAS replicate 62% of 97 previously published associations for aesthetic traits in dogs and find 30 new associations (SI Appendix, Table S3 and Dataset S3), confirming that we have sufficient power to find associations with highly heritable traits at our current sample sizes. It also illustrates the power of owner surveys for finding loci associated with aesthetic traits using simple survey questions (Fig. 2). We find an association with black coat color in the gene EPB41L1 (chr24:25568410; 3.91 × 10^–10; beta=0.110), which is also associated with hair color in humans (87), and an association with red, brown, or tan coat color at the gene MFSD12 (chr20:56248427; P = 9.39 × 10^–22; beta = –0.165), which was previously linked to pigment intensity in both dogs (88) and humans (89). We found eight associations for behavioral traits not previously reported in dogs (SI Appendix, Table S6). The strongest association, to how affectionate owners say their dog is, is in the gene SH3BP4, which may be associated with Alzheimer’s disease in humans (90) (Fig. 3A).

Fig. 2.

GWAS in dogs using survey phenotypes finds known and novel associations with aesthetic traits. On the Darwin’s Ark survey platform, the word “DOG” in the question is replaced with the dog’s name to personalize the experience. Rather than requesting exact measurements, we collect estimates for (A) height relative to human height and (B) fur length relative to human finger length. We use visual aids to ask about (C) leg length and (D) ear shape. A question about coat color finds associations for (E) black and (F) red, liver, brown, or tan.

Fig. 3.

Few significant associations with behavior traits. (A) The strongest associations are to how affectionate a dog is, which is a moderately heritable trait (h²_SNP=0.31 ± 0.082 SE). (B) Some phenotypes like Q57 “DOG is easy to train” are highly heritable (h²_SNP = 0.48 ± 0.07 SE) but have no significant associations at current sample sizes, suggesting a polygenic architecture.

We did not replicate any of the associations reported in the breed-average studies. Out of the 135 candidate behavior SNPs in the GWAS, none had significant associations with any behavioral trait (SI Appendix, Table S7 and Fig. 1D). We found four significant associations of candidate behavior SNPs with aesthetic traits. None of the candidate behavior SNPs overlapped regions associated with behavioral traits, but 17 overlapped regions associated with an aesthetic trait, including nine associated with coat color, eight associated with ear shape, six associated with fur length, and six associated with height and leg length (SI Appendix, Table S6).

Behavioral traits are less heritable than aesthetic traits.

We confirmed that behavioral traits have lower heritability than aesthetic traits, suggesting that even with much larger GWAS, attempts to predict the behavioral characteristics of individual dogs using genetic information alone will have limited accuracy for many traits (3). Heritability for eight behavioral factors and 34 behavioral questions ranged from h²_snp = 0.08 to 0.48 (mean = 0.28 ± 0.089 SD) (Fig. 1E) (SI Appendix, Table S8). At current sample sizes, our most heritable behavior question (Q107 “Dog is easy to train”; h²_SNP = 0.48 ± 0.07SD) had no significant associated peaks (Fig. 3B). In contrast, aesthetic traits are highly heritable, ranging from h²_snp = 0.48 to 1.00 (mean = 0.68 ± 0.24 SD).

Candidate Behavior SNPs Are Correlated With Breed Ancestry.

The candidate behavior SNPs from the breed-average studies are unusually differentiated between breeds, which may explain why they appeared associated with traits in breed-average comparisons. Even in studies using individual-level data, population stratification, which is when allele frequencies and traits covary due to shared ancestry rather than causality, can produce spurious associations at highly differentiated markers (91). Many of the candidate behavior SNPs have unusually high differentiation, indicating that a large proportion of their genetic variation is structured by breed (Fig. 4A). Among breeds represented by more than 20 single-breed dogs in Darwin’s Ark, Wright’s F_st for the 135 candidate behavior SNPs in the GWAS averages 0.19 ± 0.13 SD and 23 SNPs have F_st > 0.25 (SI Appendix, Table S2) (92).

Fig. 4.

Evidence of population stratification at candidate behavior SNPs. (A) Candidate SNPs have significantly higher differentiation between breeds compared to a set of random SNPs. Violin plots show the density of F_ST values, with horizontal lines indicating the median. Jittered points are overlaid for candidate SNPs. Significance measured with a one-sided Wilcoxon rank-sum test. (B) SNP-level correlations between genotype and breed ancestry for candidate SNPs correlated with ancestry in at least 5 breeds (|r| > 0). Each boxplot summarizes correlations across breeds for one SNP (gray boxes) and shows the median (horizontal line), interquartile range (box), and whiskers extending to the smallest and largest values within 1.5 × IQR of the lower and upper quartiles. Breeds beyond the whiskers are plotted as outliers. (C) Heatmap of the 10 SNPs most strongly associated with breed height (|r| with height > 0.3, P < 0.05) and their correlations with the ancestry proportion of 10 dog breeds spanning the height spectrum. Fill color and overlaid text indicate the Pearson correlation between SNP genotype and breed ancestry. SNPs are ordered by correlation with breed height; breeds are ordered by height. Labels include breed name and approximate median height, as well as gene names for SNPs near known size-associated loci (e.g., IGF1, FGF4).

Many of the candidate SNPs are also strongly correlated with breed ancestry, which is further evidence of population stratification. We measured the correlation between genotype and breed ancestry estimates for the 70 candidate behavior SNPs overlapping with Darwin’s Ark (Dataset S4). For each of 101 breeds, we calculated the correlation between SNP genotype and the proportion of ancestry from that breed in each dog, including both mixed-breed and single-breed dogs. Two-thirds (47 SNPs) are correlated with ancestry from at least one breed (|R_spearman| > 0.1), and 30% (21 SNPs) are correlated with ancestry from at least five breeds (Fig. 4B). At a subset of SNPs, the correlation of allele frequency with breed varies with breed standard height in an almost textbook illustration of population stratification (Fig. 4C). This is particularly evident at SNPs near genes previously associated with height or leg length such as IGF1, IGF2BP2, HMGA2, and the Fgf4 retrogene on chromosome 18 (30, 72, 74, 75, 77).

Breed-Average Studies Were Confounded by Population Structure.

The most likely explanation for the strong associations observed in breed-average studies is confounding due to population structure. None of the candidate SNPs were significantly associated with behavior phenotypes in GWAS in a genetically diverse population of pet dogs. Furthermore, we detected very few significant associations for any other behavioral traits, consistent with our sample size of ~3,000 dogs being underpowered for complex trait GWAS. Consequently, developing meaningful polygenic risk scores, which would require substantially larger cohorts and independent validation datasets, remains unfeasible.

To assess whether these candidate SNPs may have initially been flagged in breed-average studies because of population structure, we took a straightforward modeling approach. For each of the 38 behavioral traits and four aesthetic traits, we tested the predictive value of each candidate SNP using a likelihood ratio test comparing a full regression model (including the SNP) to a reduced model (excluding the SNP). To evaluate whether associations were mediated by other factors, we included covariates such as sex (including spay/neuter status), age, body size (height and leg length), and breed ancestry. We excluded two SNPs significantly associated with aesthetic traits not included in this model from the behavior prediction analysis. For the analysis of body size traits (height and leg length), the corresponding trait was excluded from the covariate set. All dogs were included in the analysis (N = 3,277), and P-values were adjusted for multiple testing using the Benjamini–Hochberg procedure to control the false discovery rate (FDR).

We validated our modeling approach by testing four SNPs previously implicated in aesthetic traits, each of which was also significantly associated with the same trait in our GWAS. All four variants showed strong predictive effects on their corresponding traits even after adjusting for covariates (p_FDR = 2.4 × 10⁻⁵⁵, 3.1 × 10⁻²⁷, 2.8 × 10⁻¹⁷, and 2.7 × 10⁻¹¹) (Fig. 5A). When we extended this analysis to all SNP/trait pairs, SNPs previously associated with specific aesthetic traits were significantly more predictive of those same traits (N = 19; median p_FDR = 0.031; range = 2.4 × 10⁻⁵⁵ − 0.93) than of other aesthetic traits (N = 261; median p_FDR = 0.75; range = 4.8 × 10⁻⁷ − 1.0) (Fig. 5B).

Fig. 5.

Predictive power of candidate behavior SNPs diminishes after accounting for size and breed ancestry. (A) Adjusted P-values (−log₁₀p_FDR) from predictive models for both aesthetic (Top) and behavioral (Bottom) traits. Each point represents one SNP–trait pair. Some of the SNPs improved predictions of aesthetic traits (green) but not of behavioral traits (orange). Filled diamonds are the four validation SNPs, and circles are all other SNPs (B) Boxplots comparing the predictive power of SNPs previously associated with specific traits versus all other SNP/trait pairs, shown separately for aesthetic (Left) and behavioral (Right) traits. (C) Violin plots showing the distribution of −log₁₀-transformed P-values across increasingly stringent covariate models for behavioral traits and (D) aesthetic traits. Lines connect values for the same SNP/trait pair across models. In all panels, the dashed red line indicates the significance threshold. In panels B–D, differences between groups were assessed using a linear mixed-effects model of −log₁₀-transformed p_FDR values, with SNP included as a random intercept to account for nonindependence. Reported values represent the estimated fixed effect (β) and associated P-value.

By contrast, candidate behavior-associated SNPs showed little such specificity. After adjusting for covariates, they were similarly predictive of the previously associated behavioral trait (N = 481; median p_FDR = 0.82; range = 0.064 to 1.0) and of other behavioral traits (N = 2,383; median p_FDR = 0.87; range = 0.031 to 1.0) (Fig. 5B). Only one SNP (chr14:15006461) showed a marginally significant effect on a behavioral trait (“Easy to train”; p_FDR = 0.03), and this was not the same trait reported in earlier studies.

The strong predictive effects of candidate SNPs on behavioral traits reported in previous studies are likely due to confounding from population stratification and artificial selection on aesthetic traits. When body size and breed are not included in our model, many candidate SNPs appear to replicate previous associations: 18 SNPs showed a significant predictive effect on at least one previously reported behavioral trait (Fig. 5C), and 12 SNPs on a reported aesthetic trait (Fig. 5D). However, after adjusting for both body size and breed ancestry, none of the SNPs remained significantly predictive of behavior. In contrast, 75% aesthetic trait associations persisted, with 9 of the 12 SNP–trait pairs still significant (Fig. 5D).

Discussion

We find no evidence that genetic variants identified through breed-average studies have utility for predicting behavior. Instead, we find that the apparent associations were likely driven by population structure and artificial selection on aesthetic traits such as body size, leg length, and ear shape. Thus, behavioral predictions based on these SNPs could reinforce misleading and potentially harmful breed stereotypes (93, 94). Our findings demonstrate a fundamental flaw in breed-average study design: Averaging phenotypes by breed and then testing for genetic predictors is inherently circular. Furthermore, breed-level comparisons lack the within-breed variance necessary to distinguish causal effects from spurious correlations.

Our results also underscore the broader challenges of developing genetic predictions for behavioral traits in dogs. While our sample sizes (~1,000 to ~3,000 dogs per trait) are smaller than typical human GWAS, they remain among the largest available for pet dogs. We show that behavioral traits consistently have lower heritability and fewer significant associations compared to aesthetic traits, highlighting behavior’s polygenic, environmentally influenced nature. Thus, much larger sample sizes are needed to detect robust associations that explain a greater proportion of trait variation (95).

Behavioral traits in dogs cannot be meaningfully predicted using a handful of candidate SNPs. Instead, more sophisticated approaches developed for complex traits are needed. Polygenic risk scores (PRS), now widely used in human medicine (26), and genomic estimated breeding values (GEBVs), established in animal breeding (96), are promising options. Both methods account for polygenicity and population structure when estimating and incorporating SNP effects across the whole genome, resulting in greater predictive power (4).

Dogs present unique challenges. The population structure of pet dogs is unlike that of any other species commonly used in genetic studies. In the Darwin’s Ark dataset alone, at least 85 breeds contribute detectable ancestry and approximately 40% of dogs carry ancestry from at least four breeds. Each breed reflects a history of differing origins, population bottlenecks, strong selection for various aesthetic traits, and ongoing inbreeding (97, 98). Some breeds, like the basenji, were derived from larger free-living dog populations (99), while others, like the great Pyrenees, trace their origins to landrace breeds shaped without human-directed breeding (100, 101). Many modern breeds were created by mixing older breeds, often leading to unexpected relatedness (11, 102). For example, the golden retriever was established in the early 1,900 s by Lord Tweedmouth who mixed Labrador retrievers, Irish setters, bloodhounds, and now-extinct breeds (103). Free-breeding village dogs without evident breed ancestry make up ~80% of the global dog population (104) and are also represented in the US pet population (16).

This complex structure introduces a major confounding factor in genetic studies of polygenic traits (15, 84, 105). Prediction accuracy is sensitive to ancestry and often decreases substantially when applied to individuals from underrepresented genetic backgrounds (91, 92). For working dog programs, where a few breeds dominate and breeding is controlled, genomic prediction with GEBVs can help select for complex traits such as behavior and health, accelerating the production of successful assistance and service dogs (93, 94). In contrast, accurately predicting trait or disease risk in pet dogs is likely to prove more challenging.

Meaningful, accurate genetic predictions for complex traits could transform veterinary medicine by identifying individuals at high risk of developing diseases like cancer (106) and helping dog breeders improve population health (96). To build accurate and generalizable predictors, large and diverse cohorts are essential. Efforts such as Darwin’s Ark (16), Dog10K (86), and the Dog Aging Project (18) are sequencing thousands of dogs. Expanding these efforts is essential for developing effective genomic tools that can enhance the health and welfare of all dogs.

Materials and Methods

Phenotypes and Genetic Data for GWAS.

Darwin’s Ark. Data were collected through Darwin’s Ark, an open-access resource for collecting owner-reported phenotypes and genetic data in dogs, as previously described (16). Briefly, dog owners were invited to participate in an online community science project (www.darwinsark.org), where they completed profile information and behavioral and physical trait surveys about their dogs.

Age. Multiple fields related to age were available across different phases of the project. For each dog, we computed a composed age value using the following logic: 1) When available, we used the auto-generated “age_norm” field, which calculates age from the dog’s date of birth; 2) If “age_norm” was unavailable (in earlier-enrolled dogs), we estimated age using the profile creation date; 3) If neither was available, we used the legacy free-text field “age,” remapping month- and week-based entries into numerical values. After applying this logic, 3,082 dogs had usable age values.

Body size. Body height was assessed using a categorical survey question (question #121; SI Appendix, Fig. S2), in which owners indicated whether their dog was ankle-high, calf-high, knee-high, thigh-high, or hip-high. A total of 285 dogs did not have responses to this question. For analyses requiring height information, we first calculated the average height for each breed using genotyped dogs with responses to question #121. We then imputed height for the dogs without height phenotypes based on the breed contributing the highest proportion to their inferred genetic ancestry. One dog lacked both ancestry information and a height response and was excluded from height-based analyses.

Survey data analysis. To prepare the data for input to GWAS, we scaled the survey questions; factor scores were already normalized. We calculated factor scores by applying the factor pattern loadings derived from our previous varimax rotation factor analysis (16), to the survey responses to generate a factor score for our updated set of dogs. The quantitative covariates used in the GWAS and heritability calculations are age and height, while the discrete covariates are sex and neuter status. For the height GWAS, the only quantitative covariate used was age.

Genetic Data.

Dogs were sequenced using Illumina NovaSeq 6000 with low-coverage sequencing with a mean coverage of ~0.82X. A total of 3,285 FASTQ files (available on SRA under PRJNA675863) were aligned to the canFam4 (UU_Cfam_GSD_1.0 + ROSY) genome assembly using NVIDIA Clara Parabrick’s (version 4.0) fq2bam wrapper for BWA-MEM on the Sol supercomputer of Arizona State University (107). The parameter --bwa-options = “-K 100000000 -Y” was included in the pbrun fq2bam command as recommended by the Dog10K consortium’s genome alignment code (https://github.com/jmkidd/dogmap). For dogs with multiple sequencing runs, sequence data were merged to improve coverage, with the exception of dogs 9881 and 5954, where higher-coverage (4×) data were retained and lower-coverage (0.2×) data were excluded. We excluded 16 dogs with mean sequencing depth < 0.1x. After filtering, 3,277 samples remained.

We performed genotype imputation with GLIMPSE2 (108). Briefly, the Dog10K phased imputation reference panel of 1,929 dogs was used to create chunks and binary reference panels for each autosome (chromosomes 1 to 38). The 3,277 binary alignment map files (BAMs) were provided via a list to GLIMPSE2_phase for phasing and imputation. The resulting individual binary variant call format files (BCFs) per chunk were ligated together into a single BCF file containing the imputed genotype calls for all dogs.

Heritability and GWAS.

Workflow. We developed a reproducible and scalable Nextflow workflow for heritability estimation and GWAS using PLINK (v1.90b6.21 and v2.00a5LM) (92) and Genome-wide Complex Trait Analysis (GCTA v1.94.1) (109). This framework, packaged in a Docker container image, enables consistent execution across computing environments and parallelization across multiple phenotypes. The workflow follows the methods we used previously (16) and includes quality control filtering, estimation of SNP-based heritability using restricted maximum likelihood (REML), and GWAS using a mixed linear model with a leave-one-chromosome-out (LOCO) approach. We note that previous work has proposed that GCTA-based heritability estimates are inflated by population stratification (109), but this is disputed by the software’s authors (110).

Quality control. After filtering using parameters defined in the Nextflow workflow script, the dataset included genotypes for 10,113,058 autosomal biallelic SNPs in 3,277 dogs. SNPs were excluded if they had a minor allele frequency (MAF) below 0.01, were missing in more than 5% of individuals, or showed extreme deviation from Hardy–Weinberg equilibrium (P < 1 × 10⁻²⁰ in the exact test with mid-p adjustment and excess heterozygosity; --hwe 1e-20 midp keep-fewhet).

Heritability and GWAS. For LD-adjusted heritability estimation, we calculated linkage disequilibrium (LD) scores in 250-kb windows (params.ld_window_kb = 250) using a block size of 10,000 kb with 5,000 kb overlap between blocks. These LD scores were then used to construct four LD-stratified GRMs for use in GCTA’s GREML-LDMS (LD score–stratified genomic REML) heritability analysis (111). We performed GWAS on autosomal chromosomes using the MLMA-LOCO method in GCTA (112). To define the associated regions, we used linkage-disequilibrium-based clumping. We defined clumps as index SNPs (SNPs with P-values < 1 × 10⁻⁶) and any other SNPs in linkage disequilibrium (r² > 0.2) and within 250 kb of an index SNP.

Ancestry estimation and population differentiation. Breed ancestry estimates for each dog were generated using ADMIXTURE (113), as previously described (16). Our dataset of 2,454 mixed-breed dogs excludes 504 dogs whose owners self-reported that they were “purebred” and 304 dogs with over 85% ancestry from a single breed. Our dataset of single-breed dogs included all dogs excluded from the mixed-breed set. We validated these thresholds previously (16). For all single-breed dogs from the ten breeds with ≥20 dogs (297 dogs total), we computed F_ST using plink1.9 for the 135 candidate SNPs and 480,988 random sampled SNPs. (SI Appendix, Table S2). The 10 breeds were Labrador retriever, golden retriever, American pit bull terrier, border collie, miniature schnauzer, Australian shepherd, dachshund, Yorkshire Terrier, German shepherd, and boxer.

Genotype correlation with breed. For each candidate SNP, we computed the Spearman correlation coefficient with breed ancestry in python using the pandas library corr method (114) for all 101 breeds (Dataset S4). We used breed ancestry estimates for the five most highly correlated breed ancestries in our prediction analysis as covariates.

Effect on Prediction.

For each of the candidate SNPs in the prediction analysis (SI Appendix, Table S2), we tested whether the genetic variant had an effect on each trait using a likelihood ratio test and the imputed genotype dosages for each SNP for each dog. We used ordinary least squares regression, implemented in the python package statstmodels in statsmodels.formula.api.ols, to fit a model predicting the phenotype from factors. We tested for significance within both the dataset of all dogs, as well as within just the mixed-breed dog dataset (Dataset S2).

First, we fit a “full” model, predicting the phenotype using the SNP of interest, in addition to covariates, including age, sex, spay/neuter status, height, leg length (Q246), and the top five breed ancestries for the SNP of interest. For testing the effect on phenotypes of height or leg length, we excluded the corresponding phenotype from the covariates used in prediction. We then fit a “reduced” model using the same covariates but without the SNP of interest. When using height as a covariate, we included the imputed height phenotypes to maximize our sample size. However, when testing for an effect on height predictions, we used only the reported height phenotypes. We excluded from the behavior prediction analysis two SNPs that were significantly associated in our GWAS with aesthetic traits not included as covariates in the full model. SNP chr13:8604988 was associated with “Q127 Fur length” (P = 6.8 × 10⁻¹¹) and “Q128 Fur texture” (P = 3.6 × 10⁻⁸) and SNP chr24:23910741 was associated with “Q243 Coat color” (P = 5.8 × 10⁻²⁵) (Dataset S3).

To determine significance of the SNP in predicting the phenotype, we used a likelihood ratio statistic of −2*(reduced_ll−full_ll), where reduced_ll and full_ll correspond to the log likelihood of each corresponding model. We determined significance of the likelihood ratio statistic by computing a P-value using the chi squared (scipy.statst.chi2.sf) with one degree of freedom.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (TXT)

Dataset S02 (XLSX)

Dataset S03 (XLSX)

Dataset S04 (XLSX)

Data, Materials, and Software Availability

Genetic analysis files; FASTQ files for all genetic information; code for curating the phenotypes and predicting SNP effects on phenotype; pipeline for GWAS and heritability; software packaged in a Docker container data have been deposited in Dryad; SRA; GitHub; Docker Hub [https://doi.org/10.5061/dryad.83bk3jb4r (44); PRJNA675863 (115); https://github.com/broadinstitute/dog_behavioral_gwas_paper (116); https://github.com/VistaSohrab/dog-gwas-heritability-nextflow/ (117); https://hub.docker.com/r/vistasohrab/terra-dogagingproject-gwas (118)]. All study data are included in the article and/or supporting information.