Accuracy of local ancestry inference and its impact on genomic prediction in admixed dairy cattle populations

Accurate local ancestry (LA) inference is important for genomic evaluations in admixed dairy cattle. This study had 2 main objectives: to compare the performance of 3 LA inference software (AllOr version 7.3, BreedOrigin, and ChromoPainter) and to evaluate the impact of LA inference errors on genomic prediction accuracy. Two simulated admixed populations were used: a structured DairyCross population created through 3-way rotational crossbreeding, and a Red Dairy cattle (RDC) population representing historical admixture among multiple breeds. For LA software evaluation, true breed-of-origin labels were available to benchmark accuracy in LA inference. To assess robustness of genomic prediction, we varied heritability (h² = 0.1 or 0.4), QTL effect correlation (ρ = 1.0 or 0.3), and the proportion of LA assignment errors at 3 levels (1%, 5%, and 10%) across the genome, referring to the fraction of alleles misassigned. Two prediction methods that incorporate LA information were tested: (1) the breed-of-origin method (BOM), which uses purebred phenotypes, and (2) the breed-of-origin of alleles method (BOA), which uses both purebred and crossbred phenotypes, to estimate SNP effects. We also employed a joint SNP-BLUP method (Joint) as a baseline prediction method assuming the same SNP effects across breeds. Results based on simulated data showed that ChromoPainter achieved the highest LA inference accuracy (RDC 97.98%, DairyCross 99.93%) among the software but required substantially more computing time than acceptable in practice. BreedOrigin was the fastest. AllOr and BreedOrigin performed comparably well in LA inference accuracy (RDC up to 96.71% and 97.04%; DairyCross up to 99.45% and 99.81%, respectively). In genomic prediction, both BOM and BOA methods were robust to 1% to 10% LA assignment errors, with prediction accuracy declining by ∼0.5% to 5.7% (BOM) and ∼0.9% to 7.3% (BOA). The BOA consistently outperformed BOM (at h² = 0.1, +15.1%-18.0%; at h² = 0.4, ∼+8.8%), especially when ρ = 0.3, reflecting BOA's use of crossbred phenotypes in the training set. However, BOA did not always exceed Joint: under ρ = 0.1, Joint was similar to or slightly better than BOA in prediction accuracy (BOA 0.754-0.804 vs. Joint 0.799-0.805 across 0-10% LA assignment error), whereas under ρ = 0.3, BOA was generally higher. Overall, the results provide software guidelines for LA inference and support LA-informed prediction, showing that the performance depends on genetic architecture and is only slightly affected by moderate LA inference errors.