CanVAS: A Harmonized and Imputed Canine Variant Atlas1

This paper introduces CanVAS, a harmonized and imputed canine variant atlas that integrates 15 diverse datasets into a single resource of over 15,000 dogs and 9.7 million variants on the CanFam4 reference assembly, thereby overcoming previous data incompatibilities to facilitate large-scale genetic studies of complex diseases in dogs.

The Gist

Imagine you are trying to solve a massive jigsaw puzzle to understand why some dogs get sick and others don't. You have thousands of puzzle pieces, but here's the catch: they all come from different boxes. Some pieces are printed in English, some in French, some are upside down, and some are labeled with numbers while others use letters. If you try to force them together, they just don't fit.

This is exactly the problem scientists faced with dog genetics. For years, researchers studied dogs using different tools, different maps, and different rules. This meant they couldn't combine their work to get a bigger, clearer picture.

Enter CanVAS.

Think of CanVAS (Canine Variant Atlas and Summary) as the ultimate "translation and organization" tool that finally makes all these puzzle pieces fit together. Here is how it works, broken down into simple concepts:

1. The Great Cleanup (Harmonization)

Before CanVAS, if one lab said a gene was at "Position 100" and another said "Position 200," they couldn't talk to each other.

• The Analogy: Imagine trying to build a house where one architect uses blueprints in meters, another in feet, and a third uses a map where North is actually South.
• What CanVAS did: The researchers took 15 different studies (involving over 15,000 dogs from 375+ breeds, plus wolves and wild dogs) and forced them all to speak the same language. They:
  • Fixed the Maps: They updated everyone to the newest, most accurate map of the dog genome (called CanFam4).
  • Flipped the Switch: Some studies had their genetic data "upside down" (like a mirror image). CanVAS flipped them back to the right side.
  • Standardized the Labels: They made sure every genetic marker had the same name, regardless of which machine originally measured it.

2. The Magic Guessing Game (Imputation)

Once the data was cleaned up, they had a solid "backbone" of about 77,000 genetic markers. But that's like having a low-resolution photo; you can see the dog's shape, but not the details of its fur or eyes.

• The Analogy: Imagine you have a sketch of a dog with only a few dots. You want to see the whole picture. You have a giant, super-detailed photo of a similar dog (the Dog10K reference panel, which contains full DNA sequences of nearly 2,000 dogs).
• What CanVAS did: Using a smart computer program called Beagle, they used the detailed "photo" to fill in the missing dots on the "sketch."
• The Result: They didn't just guess; they statistically predicted the missing genetic information with high confidence. They turned 77,000 markers into 9.7 million markers. This is like turning a blurry black-and-white sketch into a high-definition, full-color movie.

3. Why This Matters

Now that we have this massive, unified database, scientists can do things they couldn't do before:

• Find Rare Diseases: Before, rare genetic glitches were hidden in the noise. Now, with 9.7 million markers, they can spot the tiny, rare changes that cause specific diseases.
• Compare Breeds: They can finally compare a Golden Retriever to a Chihuahua or a Wolf on the exact same playing field to see how their DNA differs.
• Check Family Trees: The paper used this data to measure "inbreeding" (how closely related dogs are). They found that some breeds, like the New Guinea Singing Dog, are very closely related (high inbreeding), while village dogs are very diverse.

The "Fine Print" (What to Watch Out For)

The paper is honest about a few glitches. On two specific chromosomes (number 27 and 32), the "map update" was tricky, so the computer guesses on those specific sections aren't as sharp as the rest. Scientists are advised to be extra careful when looking at those two specific areas.

The Bottom Line

CanVAS is like building a single, giant library where every dog genetic study ever done is now shelved in the same order, written in the same language, and filled in with missing pages. It turns a chaotic pile of scattered notes into a powerful, unified tool that will help veterinarians and researchers cure diseases and understand the amazing diversity of our best friends.

Technical Summary

1. Problem Statement

The domestic dog (Canis lupus familiaris) is a premier model for genetic studies of complex diseases due to its population structure (distinct breeds acting as natural genetic isolates). However, the utility of canine genomic data has been severely limited by data fragmentation and incompatibility.

• Fragmentation: Genotype data is scattered across 15+ independent studies.
• Technical Incompatibility: These datasets suffer from:
  • Incompatible Platforms: Primarily Illumina CanineHD (170k SNPs) and Thermo Fisher Axiom (710k–1.1M SNPs).
  • Inconsistent Genome Builds: Data exists on CanFam2, CanFam3.1, and the newer CanFam4.
  • Divergent Conventions: Conflicting strand orientations (Illumina TOP/BOT vs. forward), allele coding schemes (numeric vs. nucleotide), and SNP identifier formats.
• Consequence: Researchers cannot easily merge cohorts for meta-analyses, leaving substantial statistical power unrealized. Furthermore, existing datasets lack the density required for fine-mapping causal variants without whole-genome sequencing (WGS).

2. Methodology

The authors developed CanVAS, a comprehensive pipeline to harmonize, merge, and impute canine genotype data.

A. Data Integration & Quality Control (QC)

• Source: 15 publicly available datasets (Dryad, Zenodo, institutional commons) comprising 15,451 dogs from >375 breeds, village dogs, dingoes, wolves, and coyotes.
• Per-Cohort QC: Applied PLINK v1.9 filters (missingness <2% for SNPs, <5% for samples). Notably, Hardy-Weinberg Equilibrium (HWE) filtering was omitted for multi-breed cohorts to avoid the "Wahlund effect" (which causes false positives in mixed populations).
• Harmonization Pipeline:
  1. Genome Build Unification: All data was first mapped to CanFam3.1 using probe IDs as stable keys (avoiding error-prone chain-file liftover for older builds). Finally, the entire dataset was lifted over to CanFam4 (UU_Cfam_GSD_1.0). Special handling was applied to Chromosomes 27 and 32, which underwent coordinate reversal in CanFam4.
  2. Strand & Allele Alignment: Empirical concordance analysis against the largest single-platform cohort (GRLS, Axiom) was used to identify and flip TOP/BOT strand conventions. Numeric allele coding (1/2/3/4) was converted to nucleotide codes.
  3. Merging: A shared backbone of 77,215 high-quality SNPs (present in ≥90% of cohorts) was established. Duplicate samples across cohorts were removed via Identity-by-Descent (IBD) analysis.

B. Imputation

• Reference Panel: The Dog10K panel (1,929 whole-genome sequenced canids from 321 breeds).
• Algorithm: Beagle 5.4 was used for imputation.
• Filtering: Post-imputation, variants were retained if they met DR² ≥ 0.3 (imputation quality) and MAF ≥ 0.01.

C. Validation

The resource was validated through:

• Allele concordance checks.
• Principal Component Analysis (PCA) and UMAP for population structure.
• Runs of Homozygosity (ROH) analysis to assess inbreeding patterns.
• Comparison of allele frequencies at known trait loci (e.g., IGF1, MC1R).

3. Key Contributions

• CanVAS Resource: A single, analysis-ready PLINK file set containing:
  • Typed Backbone: 77,215 SNPs across 15,451 dogs.
  • Imputed Dataset: 9.7 million variants (DR² ≥ 0.3), expanding the data density by ~125-fold.
• Standardization: A reproducible pipeline (available on GitHub) that resolves complex issues of genome builds, strand orientation, and probe ID mapping across disparate platforms.
• Metadata: Comprehensive annotations including harmonized breed names (386 canonical labels), cohort origins, and phenotype data.
• Open Access: All data is deposited on Zenodo, and all scripts are available under an MIT license.

4. Results

• Data Scale: The final resource covers 15,451 dogs, including rare populations (e.g., New Guinea Singing Dogs, Chinese indigenous dogs) and wild canids.
• Imputation Quality:
  • The imputation yielded 9,667,790 variants passing filters.
  • Mean DR² across autosomes was 0.55–0.62.
  • Chromosomes 27 and 32 showed reduced quality (Mean DR² = 0.08 and 0.29) due to the coordinate reversal in the CanFam4 assembly, serving as a known limitation.
• Allele Frequency Spectrum: Imputation successfully recovered the "L-shaped" site frequency spectrum expected under neutral theory, adding ~3 million rare variants (MAF < 0.05) that were absent in the typed array data.
• Population Structure:
  • PCA and UMAP clearly separated breeds (e.g., Golden Retrievers, Boxers, Bernese Mountain Dogs).
  • No systematic batch effects were detected; breeds from different cohorts co-localized.
• Inbreeding Analysis (ROH):
  • Calculated genomic inbreeding coefficients ($F_{ROH}$) for all dogs.
  • Findings: Village dogs showed the lowest inbreeding ($F_{ROH} \approx 0.04–0.08$), while purebreds like New Guinea Singing Dogs, Skye Terriers, and Bull Terriers showed high inbreeding ($F_{ROH} > 0.49$).
  • Breed groups (e.g., Setters, Working dogs) exhibited distinct ROH length distributions, reflecting different breeding histories (recent consanguinity vs. historical bottlenecks).
• Trait Locus Validation: The imputed dataset showed stronger allele frequency divergence at known causal loci (e.g., IGF1 for size, FGF5 for coat length) compared to the typed backbone, confirming the utility of the higher density for fine-mapping.

5. Significance

CanVAS represents a transformative resource for canine genomics by:

1. Enabling Meta-Analysis: It allows researchers to combine data from previously incompatible studies, significantly increasing statistical power for Genome-Wide Association Studies (GWAS).
2. Cost-Effective Sequencing: By imputing array data to whole-genome density using the Dog10K panel, it provides high-resolution genetic data without the cost of sequencing every individual.
3. Broad Applicability: The resource supports diverse research goals, including:
   • Mapping complex disease risk loci.
   • Calculating Polygenic Risk Scores (PRS) across breeds.
   • Studying evolutionary history and selection sweeps.
   • Quantifying genomic inbreeding for conservation and breeding management.
4. Reproducibility: By providing the full pipeline and scripts, the authors ensure that future datasets can be integrated into the CanVAS framework, allowing the atlas to grow over time.

Limitations: Users must exercise caution with Chromosomes 27 and 32 due to the assembly coordinate issues, and cohort should be included as a covariate in association studies to control for any residual population stratification.