An interpretable machine learning framework for dog breed inference and ancestry decomposition.

Area of Science:

• Canine genomics and population genetics
• Machine learning applications in bioinformatics

Background:

• Domesticated dogs exhibit over 300 breeds shaped by artificial selection and population bottlenecks.
• Accurate inference of dog breed from genomic data is challenging due to high dimensionality, uneven sampling, and admixture.

Purpose of the Study:

• To develop an interpretable machine learning framework for inferring dog breed labels from genome-wide single nucleotide polymorphism (SNP) data.
• To enable both purebred classification and mixed-breed ancestry inference.

Main Methods:

• Combined dimensionality reduction with a multi-output random forest model.
• Mapped genetic variation to a continuous representation of breed membership.
• Applied the framework to the Dog Aging Project (DAP) dataset (6,572 dogs, 100 breeds).

Main Results:

• Achieved 91.7% accuracy using an overlap-based metric, outperforming an ADMIXTURE-based benchmark (87.8%).
• Demonstrated that as few as 150 informative SNPs are sufficient for near-maximal predictive performance.
• Introduced a SNP importance score metric, revealing biologically relevant loci for morphology, pigmentation, and behavior.

Conclusions:

• The developed framework offers an accurate, flexible, and interpretable method for predicting dog breed ancestry.
• Potential applications include veterinary genomics, canine population genetics, and identifying loci for breed-specific phenotypes.
• Highlights the structured nature of canine genetic variation and the potential for genetic discovery.