Breeding values and index creation for health and behavior traits in Labrador Retriever guide dogs

Genomic breeding values and multi-trait selection indices have significantly advanced genetic improvement in livestock but remain underutilized in guide dog breeding. This study developed a genomically informed selection framework for a population of Labrador Retrievers by integrating health (e.g., dental, ocular, and dermatological conditions) and behavioral (e.g., trainability, distraction level, pace) traits into a "Behavior Score," "Health Score," and "Total Score" index by applying Genomic Best Linear Unbiased Prediction (GBLUP) to estimate breeding values.

Phenotypic and genotypic data were collected from 844 dogs over 26 years at The Seeing Eye guide dog school. Predictive performance was evaluated via five-fold cross-validation and correlation-based metrics. Results showed that some dentition related health traits exhibited moderate to high Area Under Receiving Operating Characteristic (AUROC) values (0.79-0.87), indicating potential for immediate use for genetic improvement. In contrast, most other health traits demonstrated weak to moderate predictive accuracy. Behavioral traits exhibited lower predictive accuracy but showed a stronger association with training success. Models were commonly unable to correctly classify individuals for binary or ordinal traits yet performed well in ranking individuals, likely due to lower heritability or strong environmental influences of traits or limitations of the dataset itself. The behavior-focused Total Score (AUROC ~0.72) outperformed health-based indices as a fixed effect in predicting breeding success despite the weaker predictive ability of individual behavioral traits. Incorporating parental scores as fixed effects modestly improved breeding values for success, indicating the importance of integrating additional data sources where available.

While these findings underscore the utility of genomic selection for guide dog breeding, they also highlight constraints stemming from small, genetically homogeneous populations and variable phenotyping. Ultimately, we provide the first usable individual and multi-trait genomic approaches to enhance both health and performance outcomes in working dog programs and a foundation to expand upon the reference population and behavioral trait assessment to improve prediction accuracy in the future.