Estimation & Prediction

Development of a Bayesian Probit mixture model for multi-ancestry risk score prediction of anterior cruciate ligament rupture in dogs

---

Abstract

Anterior cruciate ligament (ACL) rupture is one of the most common and debilitating orthopaedic diseases in dogs. Liability to this disease varies substantially among breeds, reflecting differences in allele frequencies, linkage disequilibrium, and underlying causal genetic architectures. These sources of heterogeneity pose a major barrier to developing portable polygenic risk scores that generalize across breeds of dog. To address this challenge, we developed an efficient and accurate multi-ancestry Bayesian probit mixture model to jointly estimate breed-specific and breed-shared SNP effects, allowing explicit quantification of genetic components that are conserved versus divergent between populations. We first evaluated the method using simulated multi-breed data generated under realistic demographic and selective scenarios. Binary phenotypes were produced through a probit liability model, and causal SNPs were assigned to one of ten structured across-breed mixture categories, including large (L), medium (M), small (S), and null (0) effect classes in each breed (LL, LM, LS, L0, MM, MS, M0, SS, S0, 00). The model employs a bivariate spike-and-slab prior, blocked SNP-wise updates, Dirichlet mixture weighting, and data-augmentation Gibbs sampling with fixed residual variance (var E=1). This design enables computationally scalable inference on tens of thousands of markers and delivers robust posterior estimates of SNP effects, mixture assignments, genetic variances, and across-breed polygenic risk scores. A key strength of the framework is its ability to classify variants according to their portability. Importantly, SNPs assigned to the LL mixture component represent large-effect loci that are shared across breeds. These portable LL markers can be readily mapped to functional genomic annotations including protein-coding elements, regulatory domains, conserved regions, and structural variants providing biological insight into connective-tissue mechanisms relevant to ACL pathology. We applied the model to a real multi-breed dataset consisting of 1,052 Labrador Retrievers (reference training population) and 108 Rottweilers (external test population), genotyped at 142K SNPs. Genotype data were quality controlled, standardized, and adjusted for population structure using principal components. A logistic mixed-model GWAS in Labradors was used to filter SNPs before model fitting. When trained in Labradors and evaluated in Rottweilers, the Bayesian probit mixture model achieved high predictive accuracy (AUC >0.85), demonstrating strong across-breed (ancestry) portability. This is a critical benchmark for real-world deployment of risk prediction tools in veterinary and human medicine. Posterior mixture probabilities allowed clear identification of LL SNPs as well as additional highly predictive markers with strong posterior inclusion probabilities and stable effect estimates. Overall, this work introduces a scalable and biologically interpretable multi-ancestry Bayesian probit mixture model that provides accurate across-breed risk prediction for ACL rupture. By partitioning SNP effects into structured across-breed categories and identifying portable LL variants, the framework offers a powerful approach for building equitable, generalizable, and mechanistically informative genomic prediction tools for orthopedic disease in dogs.

Keywords: 2026

How to Cite:

Momen, M. & Muir, p., (2026) “Development of a Bayesian Probit mixture model for multi-ancestry risk score prediction of anterior cruciate ligament rupture in dogs”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2294501. doi: https://doi.org/10.31274/wcgalp.24321

Rights: 1