Benchmarking Genomic Breed Discrimination in Cattle and Sheep: Discrimination, Calibration, and the Role of Population Structure

<title>Abstract</title> <p> Background. Reliable genomic breed assignment underpins livestock traceability, pedigree verification, and conservation breeding, yet most benchmarks address a single species and rank classifiers by discrimination alone. Methods. We benchmarked fourteen machine-learning and statistical classifiers for breed assignment in 20 cattle breeds ( <italic>n</italic>  = 403; <italic>Bos taurus</italic> , <italic>B. indicus</italic> , and <italic>B. javanicus</italic> lineages) and 20 <italic>Ovis aries</italic> sheep breeds ( <italic>n</italic>  = 1,458). Whole-chip SNP data underwent quality control, LD pruning, and autoencoder-based feature selection to 5,000 informative markers, and models were compared across discrimination, calibration, and generalisation using a composite weighted multi-metric ranking. Results. Top classifiers reached near-maximal accuracy in both species, but inter-model structure differed. In cattle, linear and probabilistic models — with both SVM variants — formed a statistically indistinguishable top tier, whereas boosting models showed train–test gaps consistent with overfitting under the high-dimensional, small-sample design. In sheep, inter-model variance was markedly compressed, consistent with the more homogeneous <italic>Ovis aries</italic> structure. A systematic accuracy–calibration dissociation emerged in both species: ridge regression and KRR (a GBLUP-analogous KernelRidge baseline with softmax-normalised, non-probabilistic posteriors) discriminated well but were poorly calibrated. PyTorch MLP, logistic regression, and elastic net occupied the top three positions, combining high accuracy with comparatively reliable posteriors. Residual misassignment — Braunvieh in cattle and five commercially selected European-derived sheep breeds — was shared across algorithmically distinct classifiers, indicating population-level genomic proximity rather than algorithmic limitation. Conclusions. Accurate, well-calibrated breed assignment is attainable in both species, and classification difficulty tracked population structure more than algorithm choice. As findings derive from within-sample cross-validation, extension to independent or admixed animals requires external validation. </p>