Image feature embedding with a deep learning framework improves genome-wide association studies on dog...

Area of Science:

• Genomics and Bioinformatics
• Animal Genetics
• Computational Biology

Background:

• Domestic dogs display significant morphological diversity, complicating phenotype characterization.
• Traditional phenotyping relies on manual measurements, insufficient for complex visual traits.
• Deep learning offers automated feature extraction from images for biological insights.

Purpose of the Study:

• To develop and validate a deep learning-based framework for automated phenotype extraction in dogs.
• To identify genotype-phenotype relationships using image-derived features and genome-wide association studies (GWAS).
• To discover novel candidate genes associated with canine morphological traits.

Main Methods:

• Constructed a dataset of 13,254 dog images across multiple breeds.
• Utilized ResNet and Vision Transformer (ViT) models to extract 256-dimensional image embeddings.
• Applied Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction, followed by GWAS on extracted features and genotype data.

Main Results:

• Identified 15 known genes associated with dog traits like hair length and body size.
• Discovered novel candidate genes, including EIF2S2, TRHR, and TCF25, linked to body development and hair growth.
• Validated the approach by confirming known genetic associations and revealing previously unidentified genotype-phenotype links.

Conclusions:

• The deep learning-based phenotype extraction framework is effective for identifying genotype-phenotype associations in dogs.
• This scalable approach enables population genetic studies and facilitates breeding programs in domestic animals.
• The study highlights the potential of AI in uncovering complex genetic underpinnings of morphological diversity.